How to recognize speech

Reference documentation | Package (NuGet) | Additional samples on GitHub

In this how-to guide, you learn how to use Azure AI Speech for real-time speech to text conversion. Real-time speech recognition is ideal for applications requiring immediate transcription, such as dictation, call center assistance, and captioning for live meetings.

To learn how to set up the environment for a sample application, see Quickstart: Recognize and convert speech to text.

Create a speech configuration instance

To call the Speech service by using the Speech SDK, you need to create a SpeechConfig instance. This class includes information about your subscription, like your key and associated region, endpoint, host, or authorization token.

1. Create a Speech resource in the Azure portal. Get the Speech resource key and region.
2. Create a SpeechConfig instance by using the following code. Replace YourSpeechKey and YourSpeechRegion with your Speech resource key and region.

using System;
using System.IO;
using System.Threading.Tasks;
using Microsoft.CognitiveServices.Speech;
using Microsoft.CognitiveServices.Speech.Audio;

class Program 
{
    async static Task Main(string[] args)
    {
        var speechConfig = SpeechConfig.FromSubscription("YourSpeechKey", "YourSpeechRegion");
    }
}

You can initialize SpeechConfig in a few other ways:

• Use an endpoint, and pass in a Speech service endpoint. A key or authorization token is optional.
• Use a host, and pass in a host address. A key or authorization token is optional.
• Use an authorization token with the associated region/location.

Note

Regardless of whether you're performing speech recognition, speech synthesis, translation, or intent recognition, you always create a configuration.

Recognize speech from a microphone

To recognize speech by using your device microphone, create an AudioConfig instance by using the FromDefaultMicrophoneInput() method. Then initialize the SpeechRecognizer object by passing speechConfig and audioConfig.

using System;
using System.IO;
using System.Threading.Tasks;
using Microsoft.CognitiveServices.Speech;
using Microsoft.CognitiveServices.Speech.Audio;

class Program 
{
    async static Task FromMic(SpeechConfig speechConfig)
    {
        using var audioConfig = AudioConfig.FromDefaultMicrophoneInput();
        using var speechRecognizer = new SpeechRecognizer(speechConfig, audioConfig);

        Console.WriteLine("Speak into your microphone.");
        var speechRecognitionResult = await speechRecognizer.RecognizeOnceAsync();
        Console.WriteLine($"RECOGNIZED: Text={speechRecognitionResult.Text}");
    }

    async static Task Main(string[] args)
    {
        var speechConfig = SpeechConfig.FromSubscription("YourSpeechKey", "YourSpeechRegion");
        await FromMic(speechConfig);
    }
}

If you want to use a specific audio input device, you need to specify the device ID in AudioConfig. To learn how to get the device ID, see Select an audio input device with the Speech SDK.

Recognize speech from a file

If you want to recognize speech from an audio file instead of a microphone, you still need to create an AudioConfig instance. However, you don't call FromDefaultMicrophoneInput(). You call FromWavFileInput() and pass the file path:

using System;
using System.IO;
using System.Threading.Tasks;
using Microsoft.CognitiveServices.Speech;
using Microsoft.CognitiveServices.Speech.Audio;

class Program 
{
    async static Task FromFile(SpeechConfig speechConfig)
    {
        using var audioConfig = AudioConfig.FromWavFileInput("PathToFile.wav");
        using var speechRecognizer = new SpeechRecognizer(speechConfig, audioConfig);

        var speechRecognitionResult = await speechRecognizer.RecognizeOnceAsync();
        Console.WriteLine($"RECOGNIZED: Text={speechRecognitionResult.Text}");
    }

    async static Task Main(string[] args)
    {
        var speechConfig = SpeechConfig.FromSubscription("YourSpeechKey", "YourSpeechRegion");
        await FromFile(speechConfig);
    }
}

Recognize speech from an in-memory stream

For many use cases, it's likely that your audio data comes from Azure Blob Storage, or it's otherwise already in memory as a byte[] instance or a similar raw data structure. The following example uses PushAudioInputStream to recognize speech, which is essentially an abstracted memory stream. The sample code does the following actions:

• Writes raw audio data to PushAudioInputStream by using the Write() function, which accepts a byte[] instance.
• Reads a .wav file by using FileReader for demonstration purposes. If you already have audio data in a byte[] instance, you can skip directly to writing the content to the input stream.
• The default format is 16-bit, 16-kHz mono pulse-code modulation (PCM) data. To customize the format, you can pass an AudioStreamFormat object to CreatePushStream() by using the static function AudioStreamFormat.GetWaveFormatPCM(sampleRate, (byte)bitRate, (byte)channels).

using System;
using System.IO;
using System.Threading.Tasks;
using Microsoft.CognitiveServices.Speech;
using Microsoft.CognitiveServices.Speech.Audio;

class Program 
{
    async static Task FromStream(SpeechConfig speechConfig)
    {
        var reader = new BinaryReader(File.OpenRead("PathToFile.wav"));
        using var audioConfigStream = AudioInputStream.CreatePushStream();
        using var audioConfig = AudioConfig.FromStreamInput(audioConfigStream);
        using var speechRecognizer = new SpeechRecognizer(speechConfig, audioConfig);

        byte[] readBytes;
        do
        {
            readBytes = reader.ReadBytes(1024);
            audioConfigStream.Write(readBytes, readBytes.Length);
        } while (readBytes.Length > 0);

        var speechRecognitionResult = await speechRecognizer.RecognizeOnceAsync();
        Console.WriteLine($"RECOGNIZED: Text={speechRecognitionResult.Text}");
    }

    async static Task Main(string[] args)
    {
        var speechConfig = SpeechConfig.FromSubscription("YourSpeechKey", "YourSpeechRegion");
        await FromStream(speechConfig);
    }
}

Using a push stream as input assumes that the audio data is raw PCM and skips any headers. The API still works in certain cases if the header isn't skipped. For the best results, consider implementing logic to read off the headers so that byte[] begins at the start of the audio data.

Handle errors

The previous examples only get the recognized text from the speechRecognitionResult.Text property. To handle errors and other responses, you need to write some code to handle the result. The following code evaluates the speechRecognitionResult.Reason property and:

• Prints the recognition result: ResultReason.RecognizedSpeech.
• If there's no recognition match, it informs the user: ResultReason.NoMatch.
• If an error is encountered, it prints the error message: ResultReason.Canceled.

switch (speechRecognitionResult.Reason)
{
    case ResultReason.RecognizedSpeech:
        Console.WriteLine($"RECOGNIZED: Text={speechRecognitionResult.Text}");
        break;
    case ResultReason.NoMatch:
        Console.WriteLine($"NOMATCH: Speech could not be recognized.");
        break;
    case ResultReason.Canceled:
        var cancellation = CancellationDetails.FromResult(speechRecognitionResult);
        Console.WriteLine($"CANCELED: Reason={cancellation.Reason}");

        if (cancellation.Reason == CancellationReason.Error)
        {
            Console.WriteLine($"CANCELED: ErrorCode={cancellation.ErrorCode}");
            Console.WriteLine($"CANCELED: ErrorDetails={cancellation.ErrorDetails}");
            Console.WriteLine($"CANCELED: Did you set the speech resource key and region values?");
        }
        break;
}

Use continuous recognition

The previous examples use single-shot recognition, which recognizes a single utterance. The end of a single utterance is determined by listening for silence at the end or until a maximum of 15 seconds of audio is processed.

In contrast, you use continuous recognition when you want to control when to stop recognizing. It requires you to subscribe to the Recognizing, Recognized, and Canceled events to get the recognition results. To stop recognition, you must call StopContinuousRecognitionAsync. Here's an example of how continuous recognition is performed on an audio input file.

Start by defining the input and initializing SpeechRecognizer:

using var audioConfig = AudioConfig.FromWavFileInput("YourAudioFile.wav");
using var speechRecognizer = new SpeechRecognizer(speechConfig, audioConfig);

Then create a TaskCompletionSource<int> instance to manage the state of speech recognition:

var stopRecognition = new TaskCompletionSource<int>();

Next, subscribe to the events that SpeechRecognizer sends:

• Recognizing: Signal for events that contain intermediate recognition results.
• Recognized: Signal for events that contain final recognition results, which indicate a successful recognition attempt.
• SessionStopped: Signal for events that indicate the end of a recognition session (operation).
• Canceled: Signal for events that contain canceled recognition results. These results indicate a recognition attempt that was canceled as a result of a direct cancelation request. Alternatively, they indicate a transport or protocol failure.

speechRecognizer.Recognizing += (s, e) =>
{
    Console.WriteLine($"RECOGNIZING: Text={e.Result.Text}");
};

speechRecognizer.Recognized += (s, e) =>
{
    if (e.Result.Reason == ResultReason.RecognizedSpeech)
    {
        Console.WriteLine($"RECOGNIZED: Text={e.Result.Text}");
    }
    else if (e.Result.Reason == ResultReason.NoMatch)
    {
        Console.WriteLine($"NOMATCH: Speech could not be recognized.");
    }
};

speechRecognizer.Canceled += (s, e) =>
{
    Console.WriteLine($"CANCELED: Reason={e.Reason}");

    if (e.Reason == CancellationReason.Error)
    {
        Console.WriteLine($"CANCELED: ErrorCode={e.ErrorCode}");
        Console.WriteLine($"CANCELED: ErrorDetails={e.ErrorDetails}");
        Console.WriteLine($"CANCELED: Did you set the speech resource key and region values?");
    }

    stopRecognition.TrySetResult(0);
};

speechRecognizer.SessionStopped += (s, e) =>
{
    Console.WriteLine("\n    Session stopped event.");
    stopRecognition.TrySetResult(0);
};

With everything set up, call StartContinuousRecognitionAsync to start recognizing:

await speechRecognizer.StartContinuousRecognitionAsync();

// Waits for completion. Use Task.WaitAny to keep the task rooted.
Task.WaitAny(new[] { stopRecognition.Task });

// Make the following call at some point to stop recognition:
// await speechRecognizer.StopContinuousRecognitionAsync();

Change the source language

A common task for speech recognition is specifying the input (or source) language. The following example shows how to change the input language to Italian. In your code, find your SpeechConfig instance and add this line directly below it:

speechConfig.SpeechRecognitionLanguage = "it-IT";

The SpeechRecognitionLanguage property expects a language-locale format string. For a list of supported locales, see Language and voice support for the Speech service.

Language identification

You can use language identification with speech to text recognition when you need to identify the language in an audio source and then transcribe it to text.

For a complete code sample, see Language identification.

Use a custom endpoint

With custom speech, you can upload your own data, test and train a custom model, compare accuracy between models, and deploy a model to a custom endpoint. The following example shows how to set a custom endpoint.

var speechConfig = SpeechConfig.FromSubscription("YourSubscriptionKey", "YourServiceRegion");
speechConfig.EndpointId = "YourEndpointId";
var speechRecognizer = new SpeechRecognizer(speechConfig);

Run and use a container

Speech containers provide websocket-based query endpoint APIs that are accessed through the Speech SDK and Speech CLI. By default, the Speech SDK and Speech CLI use the public Speech service. To use the container, you need to change the initialization method. Use a container host URL instead of key and region.

For more information about containers, see Host URLs in Install and run Speech containers with Docker.

Change how silence is handled

If a user speaks faster or slower than usual, the default behaviors for nonspeech silence in input audio might not result in what you expect. Common problems with silence handling include:

• Fast-speech that chains many sentences together into a single recognition result, instead of breaking sentences into individual results.
• Slow speech that separates parts of a single sentence into multiple results.
• A single-shot recognition that ends too quickly while waiting for speech to begin.

These problems can be addressed by setting one of two timeout properties on the SpeechConfig instance used to create a SpeechRecognizer:

• Segmentation silence timeout adjusts how much nonspeech audio is allowed within a phrase that's currently being spoken before that phrase is considered "done."
  • Higher values generally make results longer and allow longer pauses from the speaker within a phrase but make results take longer to arrive. They can also combine separate phrases into a single result when set too high.
  • Lower values generally make results shorter and ensure more prompt and frequent breaks between phrases, but can also cause single phrases to separate into multiple results when set too low.
  • This timeout can be set to integer values between 100 and 5000, in milliseconds, with 500 a typical default.
• Initial silence timeout adjusts how much nonspeech audio is allowed before a phrase before the recognition attempt ends in a "no match" result.
  • Higher values give speakers more time to react and start speaking, but can also result in slow responsiveness when nothing is spoken.
  • Lower values ensure a prompt "no match" for faster user experience and more controlled audio handling, but might cut a speaker off too quickly when set too low.
  • Because continuous recognition generates many results, this value determines how often "no match" results arrive but doesn't otherwise affect the content of recognition results.
  • This timeout can be set to any non-negative integer value, in milliseconds, or set to 0 to disable it entirely. 5000 is a typical default for single-shot recognition while 15000 is a typical default for continuous recognition.

Since there are tradeoffs when modifying these timeouts, you should only change the settings when you have a problem related to silence handling. Default values optimally handle most spoken audio and only uncommon scenarios should encounter problems.

Example: Users speaking a serial number like "ABC-123-4567" might pause between character groups long enough for the serial number to be broken into multiple results. In this case, try a higher value like 2000 milliseconds for the segmentation silence timeout:

speechConfig.SetProperty(PropertyId.Speech_SegmentationSilenceTimeoutMs, "2000");

Example: A recorded presenter's speech might be fast enough that several sentences in a row get combined, with large recognition results only arriving once or twice per minute. In this case, set the segmentation silence timeout to a lower value like 300 ms:

speechConfig.SetProperty(PropertyId.Speech_SegmentationSilenceTimeoutMs, "300");

Example: A single-shot recognition asking a speaker to find and read a serial number ends too quickly while the number is being found. In this case, try a longer initial silence timeout like 10,000 ms:

speechConfig.SetProperty(PropertyId.SpeechServiceConnection_InitialSilenceTimeoutMs, "10000");

Semantic segmentation

Semantic segmentation is a speech recognition segmentation strategy that's designed to mitigate issues associated with silence-based segmentation:

• Under-segmentation: When users speak for a long time without pauses, they can see a long sequence of text without breaks ("wall of text"), which severely degrades their readability experience.
• Over-segmentation: When a user pauses for a short time, the silence detection mechanism can segment incorrectly.

Instead of relying on silence timeouts, semantic segmentation segments and returns final results when it detects sentence-ending punctuation (such as '.' or '?'). This improves the user experience with higher-quality, semantically complete segments and prevents long intermediate results.

To use semantic segmentation, you need to set the following property on the SpeechConfig instance used to create a SpeechRecognizer:

speechConfig.SetProperty(PropertyId.Speech_SegmentationStrategy, "Semantic");

Some of the limitations of semantic segmentation are as follows:

• You need the Speech SDK version 1.41 or later to use semantic segmentation.
• Semantic segmentation is only intended for use in continuous recognition. This includes scenarios such as transcription and captioning. It shouldn't be used in the single recognition and dictation mode.
• Semantic segmentation isn't available for all languages and locales. Currently, semantic segmentation is only available for English (en) locales such as en-US, en-GB, en-IN, and en-AU.
• Semantic segmentation doesn't yet support confidence scores and NBest lists. As such, we don't recommend semantic segmentation if you're using confidence scores or NBest lists.

Reference documentation | Package (NuGet) | Additional samples on GitHub

In this how-to guide, you learn how to use Azure AI Speech for real-time speech to text conversion. Real-time speech recognition is ideal for applications requiring immediate transcription, such as dictation, call center assistance, and captioning for live meetings.

To learn how to set up the environment for a sample application, see Quickstart: Recognize and convert speech to text.

Create a speech configuration instance

To call the Speech service using the Speech SDK, you need to create a SpeechConfig instance. This class includes information about your subscription, like your key and associated region, endpoint, host, or authorization token.

1. Create a Speech resource in the Azure portal. Get the Speech resource key and region.
2. Create a SpeechConfig instance by using the following code. Replace YourSpeechKey and YourSpeechRegion with your Speech resource key and region.

using namespace std;
using namespace Microsoft::CognitiveServices::Speech;

auto speechConfig = SpeechConfig::FromSubscription("YourSpeechKey", "YourSpeechRegion");

You can initialize SpeechConfig in a few other ways:

• Use an endpoint, and pass in a Speech service endpoint. A key or authorization token is optional.
• Use a host, and pass in a host address. A key or authorization token is optional.
• Use an authorization token with the associated region/location.

Note

Regardless of whether you're performing speech recognition, speech synthesis, translation, or intent recognition, you always create a configuration.

Recognize speech from a microphone

To recognize speech by using your device microphone, create an AudioConfig instance by using the FromDefaultMicrophoneInput() member function. Then initialize theSpeechRecognizer object by passing audioConfig and config.

using namespace Microsoft::CognitiveServices::Speech::Audio;

auto audioConfig = AudioConfig::FromDefaultMicrophoneInput();
auto speechRecognizer = SpeechRecognizer::FromConfig(config, audioConfig);

cout << "Speak into your microphone." << std::endl;
auto result = speechRecognizer->RecognizeOnceAsync().get();
cout << "RECOGNIZED: Text=" << result->Text << std::endl;

If you want to use a specific audio input device, you need to specify the device ID in AudioConfig. To learn how to get the device ID, see Select an audio input device with the Speech SDK.

Recognize speech from a file

If you want to recognize speech from an audio file instead of using a microphone, you still need to create an AudioConfig instance. However, you don't call FromDefaultMicrophoneInput(). You call FromWavFileInput() and pass the file path:

using namespace Microsoft::CognitiveServices::Speech::Audio;

auto audioConfig = AudioConfig::FromWavFileInput("YourAudioFile.wav");
auto speechRecognizer = SpeechRecognizer::FromConfig(config, audioConfig);

auto result = speechRecognizer->RecognizeOnceAsync().get();
cout << "RECOGNIZED: Text=" << result->Text << std::endl;

Recognize speech by using the Recognizer class

The Recognizer class for the Speech SDK for C++ exposes a few methods that you can use for speech recognition.

Single-shot recognition

Single-shot recognition asynchronously recognizes a single utterance. The end of a single utterance is determined by listening for silence at the end or until a maximum of 15 seconds of audio is processed. Here's an example of asynchronous single-shot recognition via RecognizeOnceAsync:

auto result = speechRecognizer->RecognizeOnceAsync().get();

You need to write some code to handle the result. This sample evaluates result->Reason and:

• Prints the recognition result: ResultReason::RecognizedSpeech.
• If there's no recognition match, it informs the user: ResultReason::NoMatch.
• If an error is encountered, it prints the error message: ResultReason::Canceled.

switch (result->Reason)
{
    case ResultReason::RecognizedSpeech:
        cout << "We recognized: " << result->Text << std::endl;
        break;
    case ResultReason::NoMatch:
        cout << "NOMATCH: Speech could not be recognized." << std::endl;
        break;
    case ResultReason::Canceled:
        {
            auto cancellation = CancellationDetails::FromResult(result);
            cout << "CANCELED: Reason=" << (int)cancellation->Reason << std::endl;
    
            if (cancellation->Reason == CancellationReason::Error) {
                cout << "CANCELED: ErrorCode= " << (int)cancellation->ErrorCode << std::endl;
                cout << "CANCELED: ErrorDetails=" << cancellation->ErrorDetails << std::endl;
                cout << "CANCELED: Did you set the speech resource key and region values?" << std::endl;
            }
        }
        break;
    default:
        break;
}

Continuous recognition

Continuous recognition is a bit more involved than single-shot recognition. It requires you to subscribe to the Recognizing, Recognized, and Canceled events to get the recognition results. To stop recognition, you must call StopContinuousRecognitionAsync. Here's an example of continuous recognition performed on an audio input file.

Start by defining the input and initializing SpeechRecognizer:

auto audioConfig = AudioConfig::FromWavFileInput("YourAudioFile.wav");
auto speechRecognizer = SpeechRecognizer::FromConfig(config, audioConfig);

Next, create a variable to manage the state of speech recognition. Declare promise<void> because at the start of recognition, you can safely assume that it's not finished:

promise<void> recognitionEnd;

Next, subscribe to the events that SpeechRecognizer sends:

• Recognizing: Signal for events that contain intermediate recognition results.
• Recognized: Signal for events that contain final recognition results, which indicate a successful recognition attempt.
• SessionStopped: Signal for events that indicate the end of a recognition session (operation).
• Canceled: Signal for events that contain canceled recognition results. These results indicate a recognition attempt that was canceled as a result of a direct cancellation request. Alternatively, they indicate a transport or protocol failure.

speechRecognizer->Recognizing.Connect([](const SpeechRecognitionEventArgs& e)
    {
        cout << "Recognizing:" << e.Result->Text << std::endl;
    });

speechRecognizer->Recognized.Connect([](const SpeechRecognitionEventArgs& e)
    {
        if (e.Result->Reason == ResultReason::RecognizedSpeech)
        {
            cout << "RECOGNIZED: Text=" << e.Result->Text 
                 << " (text could not be translated)" << std::endl;
        }
        else if (e.Result->Reason == ResultReason::NoMatch)
        {
            cout << "NOMATCH: Speech could not be recognized." << std::endl;
        }
    });

speechRecognizer->Canceled.Connect([&recognitionEnd](const SpeechRecognitionCanceledEventArgs& e)
    {
        cout << "CANCELED: Reason=" << (int)e.Reason << std::endl;
        if (e.Reason == CancellationReason::Error)
        {
            cout << "CANCELED: ErrorCode=" << (int)e.ErrorCode << "\n"
                 << "CANCELED: ErrorDetails=" << e.ErrorDetails << "\n"
                 << "CANCELED: Did you set the speech resource key and region values?" << std::endl;

            recognitionEnd.set_value(); // Notify to stop recognition.
        }
    });

speechRecognizer->SessionStopped.Connect([&recognitionEnd](const SessionEventArgs& e)
    {
        cout << "Session stopped.";
        recognitionEnd.set_value(); // Notify to stop recognition.
    });

With everything set up, call StartContinuousRecognitionAsync to start recognizing:

// Starts continuous recognition. Uses StopContinuousRecognitionAsync() to stop recognition.
speechRecognizer->StartContinuousRecognitionAsync().get();

// Waits for recognition end.
recognitionEnd.get_future().get();

// Stops recognition.
speechRecognizer->StopContinuousRecognitionAsync().get();

Change the source language

A common task for speech recognition is specifying the input (or source) language. The following example shows how to change the input language to German. In your code, find your SpeechConfig instance and add this line directly below it:

speechConfig->SetSpeechRecognitionLanguage("de-DE");

SetSpeechRecognitionLanguage is a parameter that takes a string as an argument. For a list of supported locales, see Language and voice support for the Speech service.

Language identification

You can use language identification with speech to text recognition when you need to identify the language in an audio source and then transcribe it to text.

For a complete code sample, see Language identification.

Use a custom endpoint

With custom speech, you can upload your own data, test and train a custom model, compare accuracy between models, and deploy a model to a custom endpoint. The following example shows how to set a custom endpoint.

auto speechConfig = SpeechConfig::FromSubscription("YourSubscriptionKey", "YourServiceRegion");
speechConfig->SetEndpointId("YourEndpointId");
auto speechRecognizer = SpeechRecognizer::FromConfig(speechConfig);

Run and use a container

Speech containers provide websocket-based query endpoint APIs that are accessed through the Speech SDK and Speech CLI. By default, the Speech SDK and Speech CLI use the public Speech service. To use the container, you need to change the initialization method. Use a container host URL instead of key and region.

For more information about containers, see Host URLs in Install and run Speech containers with Docker.

Semantic segmentation

Semantic segmentation is a speech recognition segmentation strategy that's designed to mitigate issues associated with silence-based segmentation:

• Under-segmentation: When users speak for a long time without pauses, they can see a long sequence of text without breaks ("wall of text"), which severely degrades their readability experience.
• Over-segmentation: When a user pauses for a short time, the silence detection mechanism can segment incorrectly.

Instead of relying on silence timeouts, semantic segmentation segments and returns final results when it detects sentence-ending punctuation (such as '.' or '?'). This improves the user experience with higher-quality, semantically complete segments and prevents long intermediate results.

To use semantic segmentation, you need to set the following property on the SpeechConfig instance used to create a SpeechRecognizer:

speechConfig->SetProperty(PropertyId::Speech_SegmentationStrategy, "Semantic");

Some of the limitations of semantic segmentation are as follows:

• You need the Speech SDK version 1.41 or later to use semantic segmentation.
• Semantic segmentation is only intended for use in continuous recognition. This includes scenarios such as transcription and captioning. It shouldn't be used in the single recognition and dictation mode.
• Semantic segmentation isn't available for all languages and locales. Currently, semantic segmentation is only available for English (en) locales such as en-US, en-GB, en-IN, and en-AU.
• Semantic segmentation doesn't yet support confidence scores and NBest lists. As such, we don't recommend semantic segmentation if you're using confidence scores or NBest lists.

Reference documentation | Package (Go) | Additional samples on GitHub

In this how-to guide, you learn how to use Azure AI Speech for real-time speech to text conversion. Real-time speech recognition is ideal for applications requiring immediate transcription, such as dictation, call center assistance, and captioning for live meetings.

To learn how to set up the environment for a sample application, see Quickstart: Recognize and convert speech to text.

Recognize speech to text from a microphone

1. Create a Speech resource in the Azure portal. Get the Speech resource key and region.
2. Use the following code sample to run speech recognition from your default device microphone. Replace YourSpeechKey and YourSpeechRegion with your Speech resource key and region. Running the script starts a recognition session on your default microphone and output text:

package main

import (
	"bufio"
	"fmt"
	"os"

	"github.com/Microsoft/cognitive-services-speech-sdk-go/audio"
	"github.com/Microsoft/cognitive-services-speech-sdk-go/speech"
)

func sessionStartedHandler(event speech.SessionEventArgs) {
	defer event.Close()
	fmt.Println("Session Started (ID=", event.SessionID, ")")
}

func sessionStoppedHandler(event speech.SessionEventArgs) {
	defer event.Close()
	fmt.Println("Session Stopped (ID=", event.SessionID, ")")
}

func recognizingHandler(event speech.SpeechRecognitionEventArgs) {
	defer event.Close()
	fmt.Println("Recognizing:", event.Result.Text)
}

func recognizedHandler(event speech.SpeechRecognitionEventArgs) {
	defer event.Close()
	fmt.Println("Recognized:", event.Result.Text)
}

func cancelledHandler(event speech.SpeechRecognitionCanceledEventArgs) {
	defer event.Close()
	fmt.Println("Received a cancellation: ", event.ErrorDetails)
	fmt.Println("Did you set the speech resource key and region values?")
}

func main() {
    subscription :=  "YourSpeechKey"
    region := "YourSpeechRegion"

	audioConfig, err := audio.NewAudioConfigFromDefaultMicrophoneInput()
	if err != nil {
		fmt.Println("Got an error: ", err)
		return
	}
	defer audioConfig.Close()
	config, err := speech.NewSpeechConfigFromSubscription(subscription, region)
	if err != nil {
		fmt.Println("Got an error: ", err)
		return
	}
	defer config.Close()
	speechRecognizer, err := speech.NewSpeechRecognizerFromConfig(config, audioConfig)
	if err != nil {
		fmt.Println("Got an error: ", err)
		return
	}
	defer speechRecognizer.Close()
	speechRecognizer.SessionStarted(sessionStartedHandler)
	speechRecognizer.SessionStopped(sessionStoppedHandler)
	speechRecognizer.Recognizing(recognizingHandler)
	speechRecognizer.Recognized(recognizedHandler)
	speechRecognizer.Canceled(cancelledHandler)
	speechRecognizer.StartContinuousRecognitionAsync()
	defer speechRecognizer.StopContinuousRecognitionAsync()
	bufio.NewReader(os.Stdin).ReadBytes('\n')
}

Run the following commands to create a go.mod file that links to components hosted on GitHub:

go mod init quickstart
go get github.com/Microsoft/cognitive-services-speech-sdk-go

Now build and run the code:

go build
go run quickstart

For detailed information, see the reference content for the SpeechConfig class and the SpeechRecognizer class.

Recognize speech to text from an audio file

Use the following sample to run speech recognition from an audio file. Replace YourSpeechKey and YourSpeechRegion with your Speech resource key and region. Additionally, replace the variable file with a path to a .wav file. When you run the script, it recognizes speech from the file and output the text result:

package main

import (
	"fmt"
	"time"

	"github.com/Microsoft/cognitive-services-speech-sdk-go/audio"
	"github.com/Microsoft/cognitive-services-speech-sdk-go/speech"
)

func main() {
    subscription :=  "YourSpeechKey"
    region := "YourSpeechRegion"
    file := "path/to/file.wav"

	audioConfig, err := audio.NewAudioConfigFromWavFileInput(file)
	if err != nil {
		fmt.Println("Got an error: ", err)
		return
	}
	defer audioConfig.Close()
	config, err := speech.NewSpeechConfigFromSubscription(subscription, region)
	if err != nil {
		fmt.Println("Got an error: ", err)
		return
	}
	defer config.Close()
	speechRecognizer, err := speech.NewSpeechRecognizerFromConfig(config, audioConfig)
	if err != nil {
		fmt.Println("Got an error: ", err)
		return
	}
	defer speechRecognizer.Close()
	speechRecognizer.SessionStarted(func(event speech.SessionEventArgs) {
		defer event.Close()
		fmt.Println("Session Started (ID=", event.SessionID, ")")
	})
	speechRecognizer.SessionStopped(func(event speech.SessionEventArgs) {
		defer event.Close()
		fmt.Println("Session Stopped (ID=", event.SessionID, ")")
	})

	task := speechRecognizer.RecognizeOnceAsync()
	var outcome speech.SpeechRecognitionOutcome
	select {
	case outcome = <-task:
	case <-time.After(5 * time.Second):
		fmt.Println("Timed out")
		return
	}
	defer outcome.Close()
	if outcome.Error != nil {
		fmt.Println("Got an error: ", outcome.Error)
	}
	fmt.Println("Got a recognition!")
	fmt.Println(outcome.Result.Text)
}

Run the following commands to create a go.mod file that links to components hosted on GitHub:

go mod init quickstart
go get github.com/Microsoft/cognitive-services-speech-sdk-go

Now build and run the code:

go build
go run quickstart

For detailed information, see the reference content for the SpeechConfig class and the SpeechRecognizer class.

Run and use a container

Speech containers provide websocket-based query endpoint APIs that are accessed through the Speech SDK and Speech CLI. By default, the Speech SDK and Speech CLI use the public Speech service. To use the container, you need to change the initialization method. Use a container host URL instead of key and region.

For more information about containers, see Host URLs in Install and run Speech containers with Docker.

Reference documentation | Additional samples on GitHub

In this how-to guide, you learn how to use Azure AI Speech for real-time speech to text conversion. Real-time speech recognition is ideal for applications requiring immediate transcription, such as dictation, call center assistance, and captioning for live meetings.

To learn how to set up the environment for a sample application, see Quickstart: Recognize and convert speech to text.

Create a speech configuration instance

To call the Speech service by using the Speech SDK, you need to create a SpeechConfig instance. This class includes information about your subscription, like your key and associated region, endpoint, host, or authorization token.

1. Create a Speech resource in the Azure portal. Get the Speech resource key and region.
2. Create a SpeechConfig instance by using your Speech key and region.

import com.microsoft.cognitiveservices.speech.*;
import com.microsoft.cognitiveservices.speech.audio.AudioConfig;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.Future;

public class Program {
    public static void main(String[] args) throws InterruptedException, ExecutionException {
        SpeechConfig speechConfig = SpeechConfig.fromSubscription("<paste-your-speech-key>", "<paste-your-region>");
    }
}

You can initialize SpeechConfig in a few other ways:

• Use an endpoint, and pass in a Speech service endpoint. A key or authorization token is optional.
• Use a host, and pass in a host address. A key or authorization token is optional.
• Use an authorization token with the associated region/location.

Note

Regardless of whether you're performing speech recognition, speech synthesis, translation, or intent recognition, you'll always create a configuration.

Recognize speech from a microphone

To recognize speech by using your device microphone, create an AudioConfig instance by using the fromDefaultMicrophoneInput() method. Then initialize the SpeechRecognizer object by passing audioConfig and config.

import com.microsoft.cognitiveservices.speech.*;
import com.microsoft.cognitiveservices.speech.audio.AudioConfig;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.Future;

public class Program {
    public static void main(String[] args) throws InterruptedException, ExecutionException {
        SpeechConfig speechConfig = SpeechConfig.fromSubscription("<paste-your-speech-key>", "<paste-your-region>");
        fromMic(speechConfig);
    }

    public static void fromMic(SpeechConfig speechConfig) throws InterruptedException, ExecutionException {
        AudioConfig audioConfig = AudioConfig.fromDefaultMicrophoneInput();
        SpeechRecognizer speechRecognizer = new SpeechRecognizer(speechConfig, audioConfig);

        System.out.println("Speak into your microphone.");
        Future<SpeechRecognitionResult> task = speechRecognizer.recognizeOnceAsync();
        SpeechRecognitionResult speechRecognitionResult = task.get();
        System.out.println("RECOGNIZED: Text=" + speechRecognitionResult.getText());
    }
}

If you want to use a specific audio input device, you need to specify the device ID in AudioConfig. To learn how to get the device ID, see Select an audio input device with the Speech SDK.

Recognize speech from a file

If you want to recognize speech from an audio file instead of using a microphone, you still need to create an AudioConfig instance. However, you don't call FromDefaultMicrophoneInput(). You call fromWavFileInput() and pass the file path:

import com.microsoft.cognitiveservices.speech.*;
import com.microsoft.cognitiveservices.speech.audio.AudioConfig;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.Future;

public class Program {
    public static void main(String[] args) throws InterruptedException, ExecutionException {
        SpeechConfig speechConfig = SpeechConfig.fromSubscription("<paste-your-speech-key>", "<paste-your-region>");
        fromFile(speechConfig);
    }

    public static void fromFile(SpeechConfig speechConfig) throws InterruptedException, ExecutionException {
        AudioConfig audioConfig = AudioConfig.fromWavFileInput("YourAudioFile.wav");
        SpeechRecognizer speechRecognizer = new SpeechRecognizer(speechConfig, audioConfig);
        
        Future<SpeechRecognitionResult> task = speechRecognizer.recognizeOnceAsync();
        SpeechRecognitionResult speechRecognitionResult = task.get();
        System.out.println("RECOGNIZED: Text=" + speechRecognitionResult.getText());
    }
}

Handle errors

The previous examples only get the recognized text by using speechRecognitionResult.getText(). To handle errors and other responses, you need to write some code to handle the result. The following example evaluates speechRecognitionResult.getReason() and:

• Prints the recognition result: ResultReason.RecognizedSpeech.
• If there's no recognition match, it informs the user: ResultReason.NoMatch.
• If an error is encountered, it prints the error message: ResultReason.Canceled.

switch (speechRecognitionResult.getReason()) {
    case ResultReason.RecognizedSpeech:
        System.out.println("We recognized: " + speechRecognitionResult.getText());
        exitCode = 0;
        break;
    case ResultReason.NoMatch:
        System.out.println("NOMATCH: Speech could not be recognized.");
        break;
    case ResultReason.Canceled: {
            CancellationDetails cancellation = CancellationDetails.fromResult(speechRecognitionResult);
            System.out.println("CANCELED: Reason=" + cancellation.getReason());

            if (cancellation.getReason() == CancellationReason.Error) {
                System.out.println("CANCELED: ErrorCode=" + cancellation.getErrorCode());
                System.out.println("CANCELED: ErrorDetails=" + cancellation.getErrorDetails());
                System.out.println("CANCELED: Did you set the speech resource key and region values?");
            }
        }
        break;
}

Use continuous recognition

The previous examples use single-shot recognition, which recognizes a single utterance. The end of a single utterance is determined by listening for silence at the end or until a maximum of 15 seconds of audio is processed.

In contrast, you use continuous recognition when you want to control when to stop recognizing. It requires you to subscribe to the recognizing, recognized, and canceled events to get the recognition results. To stop recognition, you must call stopContinuousRecognitionAsync. Here's an example of how you can perform continuous recognition on an audio input file.

Start by defining the input and initializing SpeechRecognizer:

AudioConfig audioConfig = AudioConfig.fromWavFileInput("YourAudioFile.wav");
SpeechRecognizer speechRecognizer = new SpeechRecognizer(config, audioConfig);

Next, create a variable to manage the state of speech recognition. Declare a Semaphore instance at the class scope:

private static Semaphore stopTranslationWithFileSemaphore;

Next, subscribe to the events that SpeechRecognizer sends:

• recognizing: Signal for events that contain intermediate recognition results.
• recognized: Signal for events that contain final recognition results, which indicate a successful recognition attempt.
• sessionStopped: Signal for events that indicate the end of a recognition session (operation).
• canceled: Signal for events that contain canceled recognition results. These results indicate a recognition attempt that was canceled as a result of a direct cancelation request. Alternatively, they indicate a transport or protocol failure.

// First initialize the semaphore.
stopTranslationWithFileSemaphore = new Semaphore(0);

speechRecognizer.recognizing.addEventListener((s, e) -> {
    System.out.println("RECOGNIZING: Text=" + e.getResult().getText());
});

speechRecognizer.recognized.addEventListener((s, e) -> {
    if (e.getResult().getReason() == ResultReason.RecognizedSpeech) {
        System.out.println("RECOGNIZED: Text=" + e.getResult().getText());
    }
    else if (e.getResult().getReason() == ResultReason.NoMatch) {
        System.out.println("NOMATCH: Speech could not be recognized.");
    }
});

speechRecognizer.canceled.addEventListener((s, e) -> {
    System.out.println("CANCELED: Reason=" + e.getReason());

    if (e.getReason() == CancellationReason.Error) {
        System.out.println("CANCELED: ErrorCode=" + e.getErrorCode());
        System.out.println("CANCELED: ErrorDetails=" + e.getErrorDetails());
        System.out.println("CANCELED: Did you set the speech resource key and region values?");
    }

    stopTranslationWithFileSemaphore.release();
});

speechRecognizer.sessionStopped.addEventListener((s, e) -> {
    System.out.println("\n    Session stopped event.");
    stopTranslationWithFileSemaphore.release();
});

With everything set up, call startContinuousRecognitionAsync to start recognizing:

// Starts continuous recognition. Uses StopContinuousRecognitionAsync() to stop recognition.
speechRecognizer.startContinuousRecognitionAsync().get();

// Waits for completion.
stopTranslationWithFileSemaphore.acquire();

// Stops recognition.
speechRecognizer.stopContinuousRecognitionAsync().get();

Change the source language

A common task for speech recognition is specifying the input (or source) language. The following example shows how to change the input language to French. In your code, find your SpeechConfig instance, and add this line directly below it:

config.setSpeechRecognitionLanguage("fr-FR");

setSpeechRecognitionLanguage is a parameter that takes a string as an argument. Refer to the list of supported speech to text locales.

Language identification

You can use language identification with speech to text recognition when you need to identify the language in an audio source and then transcribe it to text.

For a complete code sample, see Language identification.

Use a custom endpoint

With custom speech, you can upload your own data, test and train a custom model, compare accuracy between models, and deploy a model to a custom endpoint. The following example shows how to set a custom endpoint:

SpeechConfig speechConfig = SpeechConfig.FromSubscription("YourSpeechKey", "YourServiceRegion");
speechConfig.setEndpointId("YourEndpointId");
SpeechRecognizer speechRecognizer = new SpeechRecognizer(speechConfig);

Run and use a container

Speech containers provide websocket-based query endpoint APIs that are accessed through the Speech SDK and Speech CLI. By default, the Speech SDK and Speech CLI use the public Speech service. To use the container, you need to change the initialization method. Use a container host URL instead of key and region.

For more information about containers, see Host URLs in Install and run Speech containers with Docker.

Semantic segmentation

Semantic segmentation is a speech recognition segmentation strategy that's designed to mitigate issues associated with silence-based segmentation:

• Under-segmentation: When users speak for a long time without pauses, they can see a long sequence of text without breaks ("wall of text"), which severely degrades their readability experience.
• Over-segmentation: When a user pauses for a short time, the silence detection mechanism can segment incorrectly.

Instead of relying on silence timeouts, semantic segmentation segments and returns final results when it detects sentence-ending punctuation (such as '.' or '?'). This improves the user experience with higher-quality, semantically complete segments and prevents long intermediate results.

To use semantic segmentation, you need to set the following property on the SpeechConfig instance used to create a SpeechRecognizer:

speechConfig.SetProperty(PropertyId.Speech_SegmentationStrategy, "Semantic");

Some of the limitations of semantic segmentation are as follows:

• You need the Speech SDK version 1.41 or later to use semantic segmentation.
• Semantic segmentation is only intended for use in continuous recognition. This includes scenarios such as transcription and captioning. It shouldn't be used in the single recognition and dictation mode.
• Semantic segmentation isn't available for all languages and locales. Currently, semantic segmentation is only available for English (en) locales such as en-US, en-GB, en-IN, and en-AU.
• Semantic segmentation doesn't yet support confidence scores and NBest lists. As such, we don't recommend semantic segmentation if you're using confidence scores or NBest lists.

Reference documentation | Package (npm) | Additional samples on GitHub | Library source code

In this how-to guide, you learn how to use Azure AI Speech for real-time speech to text conversion. Real-time speech recognition is ideal for applications requiring immediate transcription, such as dictation, call center assistance, and captioning for live meetings.

To learn how to set up the environment for a sample application, see Quickstart: Recognize and convert speech to text.

Create a speech configuration instance

To call the Speech service by using the Speech SDK, you need to create a SpeechConfig instance. This class includes information about your subscription, like your key and associated region, endpoint, host, or authorization token.

1. Create a Speech resource in the Azure portal. Get the Speech resource key and region.
2. Create a SpeechConfig instance by using the following code. Replace YourSpeechKey and YourSpeechRegion with your Speech resource key and region.

const speechConfig = sdk.SpeechConfig.fromSubscription("YourSpeechKey", "YourSpeechRegion");

You can initialize SpeechConfig in a few other ways:

• Use an endpoint, and pass in a Speech service endpoint. A key or authorization token is optional.
• Use a host, and pass in a host address. A key or authorization token is optional.
• Use an authorization token with the associated region/location.

Note

Regardless of whether you're performing speech recognition, speech synthesis, translation, or intent recognition, you always create a configuration.

Recognize speech from a microphone

Recognizing speech from a microphone isn't supported in Node.js. It's supported only in a browser-based JavaScript environment. For more information, see the React sample and the implementation of speech to text from a microphone on GitHub. The React sample shows design patterns for the exchange and management of authentication tokens. It also shows the capture of audio from a microphone or file for speech to text conversions.

Note

If you want to use a specific audio input device, you need to specify the device ID in AudioConfig. To learn how to get the device ID, see Select an audio input device with the Speech SDK.

Recognize speech from a file

To recognize speech from an audio file, create an AudioConfig instance by using the fromWavFileInput() method, which accepts a Buffer object. Then initialize SpeechRecognizer by passing audioConfig and speechConfig.

const fs = require('fs');
const sdk = require("microsoft-cognitiveservices-speech-sdk");
const speechConfig = sdk.SpeechConfig.fromSubscription("YourSpeechKey", "YourSpeechRegion");

function fromFile() {
    let audioConfig = sdk.AudioConfig.fromWavFileInput(fs.readFileSync("YourAudioFile.wav"));
    let speechRecognizer = new sdk.SpeechRecognizer(speechConfig, audioConfig);

    speechRecognizer.recognizeOnceAsync(result => {
        console.log(`RECOGNIZED: Text=${result.text}`);
        speechRecognizer.close();
    });
}
fromFile();

Recognize speech from an in-memory stream

For many use cases, your audio data likely comes from Azure Blob Storage. Or it's already in memory as an ArrayBuffer or a similar raw data structure. The following code:

• Creates a push stream by using createPushStream().
• Reads a .wav file by using fs.createReadStream for demonstration purposes. If you already have audio data in the ArrayBuffer, you can skip directly to writing the content to the input stream.
• Creates an audio configuration by using the push stream.

const fs = require('fs');
const sdk = require("microsoft-cognitiveservices-speech-sdk");
const speechConfig = sdk.SpeechConfig.fromSubscription("YourSpeechKey", "YourSpeechRegion");

function fromStream() {
    let pushStream = sdk.AudioInputStream.createPushStream();

    fs.createReadStream("YourAudioFile.wav").on('data', function(arrayBuffer) {
        pushStream.write(arrayBuffer.slice());
    }).on('end', function() {
        pushStream.close();
    });
 
    let audioConfig = sdk.AudioConfig.fromStreamInput(pushStream);
    let speechRecognizer = new sdk.SpeechRecognizer(speechConfig, audioConfig);
    speechRecognizer.recognizeOnceAsync(result => {
        console.log(`RECOGNIZED: Text=${result.text}`);
        speechRecognizer.close();
    });
}
fromStream();

Using a push stream as input assumes that the audio data is raw pulse-code modulation (PCM) data that skips any headers. The API still works in certain cases if the header isn't skipped. For the best results, consider implementing logic to read off the headers so that fs begins at the start of the audio data.

Handle errors

The previous examples only get the recognized text from the result.text property. To handle errors and other responses, you need to write some code to handle the result. The following code evaluates the result.reason property and:

• Prints the recognition result: ResultReason.RecognizedSpeech.
• If there's no recognition match, it informs the user: ResultReason.NoMatch.
• If an error is encountered, it prints the error message: ResultReason.Canceled.

switch (result.reason) {
    case sdk.ResultReason.RecognizedSpeech:
        console.log(`RECOGNIZED: Text=${result.text}`);
        break;
    case sdk.ResultReason.NoMatch:
        console.log("NOMATCH: Speech could not be recognized.");
        break;
    case sdk.ResultReason.Canceled:
        const cancellation = sdk.CancellationDetails.fromResult(result);
        console.log(`CANCELED: Reason=${cancellation.reason}`);

        if (cancellation.reason == sdk.CancellationReason.Error) {
            console.log(`CANCELED: ErrorCode=${cancellation.ErrorCode}`);
            console.log(`CANCELED: ErrorDetails=${cancellation.errorDetails}`);
            console.log("CANCELED: Did you set the speech resource key and region values?");
        }
        break;
    }

Use continuous recognition

The previous examples use single-shot recognition, which recognizes a single utterance. The end of a single utterance is determined by listening for silence at the end or until a maximum of 15 seconds of audio is processed.

In contrast, you can use continuous recognition when you want to control when to stop recognizing. It requires you to subscribe to the Recognizing, Recognized, and Canceled events to get the recognition results. To stop recognition, you must call [stopContinuousRecognitionAsync] (/javascript/api/microsoft-cognitiveservices-speech-sdk/speechrecognizer#microsoft-cognitiveservices-speech-sdk-speechrecognizer-stopcontinuousrecognitionasync). Here's an example of how continuous recognition is performed on an audio input file.

Start by defining the input and initializing SpeechRecognizer:

const speechRecognizer = new sdk.SpeechRecognizer(speechConfig, audioConfig);

Next, subscribe to the events sent from SpeechRecognizer:

• recognizing: Signal for events that contain intermediate recognition results.
• recognized: Signal for events that contain final recognition results, which indicate a successful recognition attempt.
• sessionStopped: Signal for events that indicate the end of a recognition session (operation).
• canceled: Signal for events that contain canceled recognition results. These results indicate a recognition attempt that was canceled as a result of a direct cancelation request. Alternatively, they indicate a transport or protocol failure.

speechRecognizer.recognizing = (s, e) => {
    console.log(`RECOGNIZING: Text=${e.result.text}`);
};

speechRecognizer.recognized = (s, e) => {
    if (e.result.reason == sdk.ResultReason.RecognizedSpeech) {
        console.log(`RECOGNIZED: Text=${e.result.text}`);
    }
    else if (e.result.reason == sdk.ResultReason.NoMatch) {
        console.log("NOMATCH: Speech could not be recognized.");
    }
};

speechRecognizer.canceled = (s, e) => {
    console.log(`CANCELED: Reason=${e.reason}`);

    if (e.reason == sdk.CancellationReason.Error) {
        console.log(`"CANCELED: ErrorCode=${e.errorCode}`);
        console.log(`"CANCELED: ErrorDetails=${e.errorDetails}`);
        console.log("CANCELED: Did you set the speech resource key and region values?");
    }

    speechRecognizer.stopContinuousRecognitionAsync();
};

speechRecognizer.sessionStopped = (s, e) => {
    console.log("\n    Session stopped event.");
    speechRecognizer.stopContinuousRecognitionAsync();
};

With everything set up, call [startContinuousRecognitionAsync] (/javascript/api/microsoft-cognitiveservices-speech-sdk/speechrecognizer#microsoft-cognitiveservices-speech-sdk-speechrecognizer-startkeywordrecognitionasync) to start recognizing:

speechRecognizer.startContinuousRecognitionAsync();

// Make the following call at some point to stop recognition:
// speechRecognizer.stopContinuousRecognitionAsync();

Change the source language

A common task for speech recognition is specifying the input (or source) language. The following example shows how to change the input language to Italian. In your code, find your SpeechConfig instance and add this line directly below it:

speechConfig.speechRecognitionLanguage = "it-IT";

The speechRecognitionLanguage property expects a language-locale format string. For a list of supported locales, see Language and voice support for the Speech service.

Language identification

You can use language identification with speech to text recognition when you need to identify the language in an audio source and then transcribe it to text.

For a complete code sample, see Language identification.

Use a custom endpoint

With custom speech, you can upload your own data, test and train a custom model, compare accuracy between models, and deploy a model to a custom endpoint. The following example shows how to set a custom endpoint.

var speechConfig = SpeechSDK.SpeechConfig.fromSubscription("YourSubscriptionKey", "YourServiceRegion");
speechConfig.endpointId = "YourEndpointId";
var speechRecognizer = new SpeechSDK.SpeechRecognizer(speechConfig);

Run and use a container

Speech containers provide websocket-based query endpoint APIs that are accessed through the Speech SDK and Speech CLI. By default, the Speech SDK and Speech CLI use the public Speech service. To use the container, you need to change the initialization method. Use a container host URL instead of key and region.

For more information about containers, see Host URLs in Install and run Speech containers with Docker.

Reference documentation | Package (download) | Additional samples on GitHub

In this how-to guide, you learn how to use Azure AI Speech for real-time speech to text conversion. Real-time speech recognition is ideal for applications requiring immediate transcription, such as dictation, call center assistance, and captioning for live meetings.

To learn how to set up the environment for a sample application, see Quickstart: Recognize and convert speech to text.

Install Speech SDK and samples

The Azure-Samples/cognitive-services-speech-sdk repository contains samples written in Objective-C for iOS and Mac. Select a link to see installation instructions for each sample:

• Recognize speech from a microphone in Objective-C on macOS
• Recognize speech in Objective-C on iOS
• More samples for Objective-C on iOS

For more information, see the Speech SDK for Objective-C reference.

Use a custom endpoint

With custom speech, you can upload your own data, test and train a custom model, compare accuracy between models, and deploy a model to a custom endpoint. The following example shows how to set a custom endpoint:

SPXSpeechConfiguration *speechConfig = [[SPXSpeechConfiguration alloc] initWithSubscription:"YourSubscriptionKey" region:"YourServiceRegion"];
speechConfig.endpointId = "YourEndpointId";
SPXSpeechRecognizer* speechRecognizer = [[SPXSpeechRecognizer alloc] init:speechConfig];

Run and use a container

Speech containers provide websocket-based query endpoint APIs that are accessed through the Speech SDK and Speech CLI. By default, the Speech SDK and Speech CLI use the public Speech service. To use the container, you need to change the initialization method. Use a container host URL instead of key and region.

For more information about containers, see Host URLs in Install and run Speech containers with Docker.

Reference documentation | Package (download) | Additional samples on GitHub

In this how-to guide, you learn how to use Azure AI Speech for real-time speech to text conversion. Real-time speech recognition is ideal for applications requiring immediate transcription, such as dictation, call center assistance, and captioning for live meetings.

To learn how to set up the environment for a sample application, see Quickstart: Recognize and convert speech to text.

Install Speech SDK and samples

The Azure-Samples/cognitive-services-speech-sdk repository contains samples written in Swift for iOS and Mac. Select a link to see installation instructions for each sample:

• Recognize speech in Swift on macOS
• Recognize speech in Swift on iOS

For more information, see the Speech SDK for Objective-C reference.

Use a custom endpoint

With custom speech, you can upload your own data, test and train a custom model, compare accuracy between models, and deploy a model to a custom endpoint. The following example shows how to set a custom endpoint:

let speechConfig = SPXSpeechConfiguration(subscription: "YourSubscriptionKey", region: "YourServiceRegion");
speechConfig.endpointId = "YourEndpointId";
let speechRecognizer = SPXSpeechRecognizer(speechConfiguration: speechConfig);

Run and use a container

Speech containers provide websocket-based query endpoint APIs that are accessed through the Speech SDK and Speech CLI. By default, the Speech SDK and Speech CLI use the public Speech service. To use the container, you need to change the initialization method. Use a container host URL instead of key and region.

For more information about containers, see Host URLs in Install and run Speech containers with Docker.

Reference documentation | Package (PyPi) | Additional samples on GitHub

In this how-to guide, you learn how to use Azure AI Speech for real-time speech to text conversion. Real-time speech recognition is ideal for applications requiring immediate transcription, such as dictation, call center assistance, and captioning for live meetings.

To learn how to set up the environment for a sample application, see Quickstart: Recognize and convert speech to text.

Create a speech configuration instance

To call the Speech service by using the Speech SDK, you need to create a SpeechConfig instance. This class includes information about your subscription, like your speech key and associated region, endpoint, host, or authorization token.

1. Create a Speech resource in the Azure portal. Get the Speech resource key and region.
2. Create a SpeechConfig instance by using the following code. Replace YourSpeechKey and YourSpeechRegion with your Speech resource key and region.

speech_config = speechsdk.SpeechConfig(subscription="YourSpeechKey", region="YourSpeechRegion")

You can initialize SpeechConfig in a few other ways:

• Use an endpoint, and pass in a Speech service endpoint. A speech key or authorization token is optional.
• Use a host, and pass in a host address. A speech key or authorization token is optional.
• Use an authorization token with the associated region/location.

Note

Regardless of whether you're performing speech recognition, speech synthesis, translation, or intent recognition, you always create a configuration.

Recognize speech from a microphone

To recognize speech by using your device microphone, create a SpeechRecognizer instance without passing AudioConfig, and then pass speech_config:

import azure.cognitiveservices.speech as speechsdk

def from_mic():
    speech_config = speechsdk.SpeechConfig(subscription="YourSpeechKey", region="YourSpeechRegion")
    speech_recognizer = speechsdk.SpeechRecognizer(speech_config=speech_config)

    print("Speak into your microphone.")
    speech_recognition_result = speech_recognizer.recognize_once_async().get()
    print(speech_recognition_result.text)

from_mic()

If you want to use a specific audio input device, you need to specify the device ID in AudioConfig, and the pass it to the SpeechRecognizer constructor's audio_config parameter. To learn how to get the device ID, see Select an audio input device with the Speech SDK.

Recognize speech from a file

If you want to recognize speech from an audio file instead of using a microphone, create an AudioConfig instance and use the filename parameter:

import azure.cognitiveservices.speech as speechsdk

def from_file():
    speech_config = speechsdk.SpeechConfig(subscription="YourSpeechKey", region="YourSpeechRegion")
    audio_config = speechsdk.AudioConfig(filename="your_file_name.wav")
    speech_recognizer = speechsdk.SpeechRecognizer(speech_config=speech_config, audio_config=audio_config)

    speech_recognition_result = speech_recognizer.recognize_once_async().get()
    print(speech_recognition_result.text)

from_file()

Handle errors

The previous examples only get the recognized text from the speech_recognition_result.text property. To handle errors and other responses, you need to write some code to handle the result. The following code evaluates the speech_recognition_result.reason property and:

• Prints the recognition result: speechsdk.ResultReason.RecognizedSpeech.
• If there's no recognition match, it informs the user: speechsdk.ResultReason.NoMatch.
• If an error is encountered, it prints the error message: speechsdk.ResultReason.Canceled.

if speech_recognition_result.reason == speechsdk.ResultReason.RecognizedSpeech:
    print("Recognized: {}".format(speech_recognition_result.text))
elif speech_recognition_result.reason == speechsdk.ResultReason.NoMatch:
    print("No speech could be recognized: {}".format(speech_recognition_result.no_match_details))
elif speech_recognition_result.reason == speechsdk.ResultReason.Canceled:
    cancellation_details = speech_recognition_result.cancellation_details
    print("Speech Recognition canceled: {}".format(cancellation_details.reason))
    if cancellation_details.reason == speechsdk.CancellationReason.Error:
        print("Error details: {}".format(cancellation_details.error_details))
        print("Did you set the speech resource key and region values?")

Use continuous recognition

The previous examples use single-shot recognition, which recognizes a single utterance. The end of a single utterance is determined by listening for silence at the end or until a maximum of 15 seconds of audio is processed.

In contrast, you use continuous recognition when you want to control when to stop recognizing. It requires you to connect to EventSignal to get the recognition results. To stop recognition, you must call stop_continuous_recognition() or stop_continuous_recognition_async(). Here's an example of how continuous recognition is performed on an audio input file.

Start by defining the input and initializing SpeechRecognizer:

audio_config = speechsdk.audio.AudioConfig(filename=weatherfilename)
speech_recognizer = speechsdk.SpeechRecognizer(speech_config=speech_config, audio_config=audio_config)

Next, create a variable to manage the state of speech recognition. Set the variable to False because at the start of recognition, you can safely assume that it's not finished:

done = False

Now, create a callback to stop continuous recognition when evt is received. Keep these points in mind:

• When evt is received, the evt message is printed.
• After evt is received, stop_continuous_recognition() is called to stop recognition.
• The recognition state is changed to True.

def stop_cb(evt):
    print('CLOSING on {}'.format(evt))
    speech_recognizer.stop_continuous_recognition()
    nonlocal done
    done = True

The following code sample shows how to connect callbacks to events sent from SpeechRecognizer. The events are:

• recognizing: Signal for events that contain intermediate recognition results.
• recognized: Signal for events that contain final recognition results, which indicate a successful recognition attempt.
• session_started: Signal for events that indicate the start of a recognition session (operation).
• session_stopped: Signal for events that indicate the end of a recognition session (operation).
• canceled: Signal for events that contain canceled recognition results. These results indicate a recognition attempt that was canceled as a result of a direct cancelation request. Alternatively, they indicate a transport or protocol failure.

speech_recognizer.recognizing.connect(lambda evt: print('RECOGNIZING: {}'.format(evt)))
speech_recognizer.recognized.connect(lambda evt: print('RECOGNIZED: {}'.format(evt)))
speech_recognizer.session_started.connect(lambda evt: print('SESSION STARTED: {}'.format(evt)))
speech_recognizer.session_stopped.connect(lambda evt: print('SESSION STOPPED {}'.format(evt)))
speech_recognizer.canceled.connect(lambda evt: print('CANCELED {}'.format(evt)))

speech_recognizer.session_stopped.connect(stop_cb)
speech_recognizer.canceled.connect(stop_cb)

With everything set up, you can call start_continuous_recognition():

speech_recognizer.start_continuous_recognition()
while not done:
    time.sleep(.5)

Change the source language

A common task for speech recognition is specifying the input (or source) language. The following example shows how to change the input language to German. In your code, find your SpeechConfig instance and add this line directly below it:

speech_config.speech_recognition_language="de-DE"

speech_recognition_language is a parameter that takes a string as an argument. For a list of supported locales, see Language and voice support for the Speech service.

Language identification

You can use language identification with Speech to text recognition when you need to identify the language in an audio source and then transcribe it to text.

For a complete code sample, see Language identification.

Use a custom endpoint

With custom speech, you can upload your own data, test and train a custom model, compare accuracy between models, and deploy a model to a custom endpoint. The following example shows how to set a custom endpoint.

speech_config = speechsdk.SpeechConfig(subscription="YourSubscriptionKey", region="YourServiceRegion")
speech_config.endpoint_id = "YourEndpointId"
speech_recognizer = speechsdk.SpeechRecognizer(speech_config=speech_config)

Run and use a container

Speech containers provide websocket-based query endpoint APIs that are accessed through the Speech SDK and Speech CLI. By default, the Speech SDK and Speech CLI use the public Speech service. To use the container, you need to change the initialization method. Use a container host URL instead of key and region.

For more information about containers, see Host URLs in Install and run Speech containers with Docker.

Semantic segmentation

Semantic segmentation is a speech recognition segmentation strategy that's designed to mitigate issues associated with silence-based segmentation:

• Under-segmentation: When users speak for a long time without pauses, they can see a long sequence of text without breaks ("wall of text"), which severely degrades their readability experience.
• Over-segmentation: When a user pauses for a short time, the silence detection mechanism can segment incorrectly.

Instead of relying on silence timeouts, semantic segmentation segments and returns final results when it detects sentence-ending punctuation (such as '.' or '?'). This improves the user experience with higher-quality, semantically complete segments and prevents long intermediate results.

To use semantic segmentation, you need to set the following property on the SpeechConfig instance used to create a SpeechRecognizer:

speech_config.set_property(speechsdk.PropertyId.Speech_SegmentationStrategy, "Semantic") 

Some of the limitations of semantic segmentation are as follows:

• You need the Speech SDK version 1.41 or later to use semantic segmentation.
• Semantic segmentation is only intended for use in continuous recognition. This includes scenarios such as transcription and captioning. It shouldn't be used in the single recognition and dictation mode.
• Semantic segmentation isn't available for all languages and locales. Currently, semantic segmentation is only available for English (en) locales such as en-US, en-GB, en-IN, and en-AU.
• Semantic segmentation doesn't yet support confidence scores and NBest lists. As such, we don't recommend semantic segmentation if you're using confidence scores or NBest lists.

Speech to text REST API reference | Speech to text REST API for short audio reference | Additional samples on GitHub

In this how-to guide, you learn how to use Azure AI Speech for real-time speech to text conversion. Real-time speech recognition is ideal for applications requiring immediate transcription, such as dictation, call center assistance, and captioning for live meetings.

To learn how to set up the environment for a sample application, see Quickstart: Recognize and convert speech to text.

Convert speech to text

At a command prompt, run the following command. Insert the following values into the command:

• Your subscription key for the Speech resource
• Your Speech service region
• The path for input audio file

curl --location --request POST 'https://INSERT_REGION_HERE.stt.speech.microsoft.com/speech/recognition/conversation/cognitiveservices/v1?language=en-US' \
--header 'Ocp-Apim-Subscription-Key: INSERT_SUBSCRIPTION_KEY_HERE' \
--header 'Content-Type: audio/wav' \
--data-binary @'INSERT_AUDIO_FILE_PATH_HERE'

You should receive a response like the following example:

{
    "RecognitionStatus": "Success",
    "DisplayText": "My voice is my passport, verify me.",
    "Offset": 6600000,
    "Duration": 32100000
}

For more information, see the Speech to text REST API reference.

In this how-to guide, you learn how to use Azure AI Speech for real-time speech to text conversion. Real-time speech recognition is ideal for applications requiring immediate transcription, such as dictation, call center assistance, and captioning for live meetings.

To learn how to set up the environment for a sample application, see Quickstart: Recognize and convert speech to text.

Recognize speech from a microphone

Plug in and turn on your PC microphone. Turn off any apps that might also use the microphone. Some computers have a built-in microphone, whereas others require configuration of a Bluetooth device.

Now you're ready to run the Speech CLI to recognize speech from your microphone. From the command line, change to the directory that contains the Speech CLI binary file. Then run the following command:

spx recognize --microphone

Note

The Speech CLI defaults to English. You can choose a different language from the speech to text table. For example, add --source de-DE to recognize German speech.

Speak into the microphone, and you can see transcription of your words into text in real time. The Speech CLI stops after a period of silence, or when you select Ctrl+C.

Recognize speech from a file

The Speech CLI can recognize speech in many file formats and natural languages. In this example, you can use any .wav file (16 kHz or 8 kHz, 16-bit, and mono PCM) that contains English speech. Or if you want a quick sample, download the file whatstheweatherlike.wav, and copy it to the same directory as the Speech CLI binary file.

Use the following command to run the Speech CLI to recognize speech found in the audio file:

spx recognize --file whatstheweatherlike.wav

Note

The Speech CLI defaults to English. You can choose a different language from the speech to text table. For example, add --source de-DE to recognize German speech.

The Speech CLI shows a text transcription of the speech on the screen.

Run and use a container

Speech containers provide websocket-based query endpoint APIs that are accessed through the Speech SDK and Speech CLI. By default, the Speech SDK and Speech CLI use the public Speech service. To use the container, you need to change the initialization method. Use a container host URL instead of key and region.

For more information about containers, see Host URLs in Install and run Speech containers with Docker.

Related content

• Try the speech to text quickstart
• Improve recognition accuracy with custom speech
• Use batch transcription