Transcription Denormalization.

Alex Cohen 10

Is there a way to "denormalize" Azure speech transcription, so it provides verbatim transcription (as close as possible, with word fillers, hesitations, repeats, etc)? I will also need word level timestamping and diarization.

I am hoping there is a simple function, like "normalize = False" for some Whisper packages.

Thanks

Amira Bedhiafi 26,261

I am not expert but I know that Azure Speech-to-Text service offers options to control the normalization of text. By default, the service may apply some normalization to improve readability. To get a more verbatim transcription, you can adjust the settings.

Azure Speech-to-Text supports word-level timestamps which you can enable to get the timing for each word in the transcription.

Diarization is the process of identifying and distinguishing between different speakers in the audio. Azure provides speaker diarization capabilities which you can enable. Here is an example, please try and tell us !

import azure.cognitiveservices.speech as speechsdk

# Initialize speech configuration
speech_key = "YourAzureSpeechKey"
service_region = "YourServiceRegion"
speech_config = speechsdk.SpeechConfig(subscription=speech_key, region=service_region)

# Set up audio configuration
audio_input = speechsdk.AudioConfig(filename="YourAudioFile.wav")

# Set up speech recognizer with desired settings
speech_recognizer = speechsdk.SpeechRecognizer(speech_config=speech_config, audio_config=audio_input)

# Enable detailed output format for word-level timestamps
speech_recognizer.recognition_format = speechsdk.SpeechRecognitionOutputFormat.Detailed

# Enable diarization
speech_recognizer.properties[speechsdk.PropertyId.SpeechServiceResponse_ProfanityOption] = "Masked"
speech_recognizer.properties[speechsdk.PropertyId.SpeechServiceResponse_ProfanityOption] = "Raw"

# Recognize speech from audio file
result = speech_recognizer.recognize_once()

# Process the result
if result.reason == speechsdk.ResultReason.RecognizedSpeech:
    detailed_result = speechsdk.SpeechRecognitionResult(result.json)
    for word in detailed_result['NBest'][0]['Words']:
        print(f"Word: {word['Word']}, Offset: {word['Offset']}, Duration: {word['Duration']}")
    print("Diarization results:", result.properties[speechsdk.PropertyId.SpeechServiceResponse_JsonResult])
else:
    print("No speech could be recognized")

Alex Cohen 10 Reputation points

2024-07-31T18:06:11.22+00:00

Thanks. Very much appreciated!

Can you specify what settings I would adjust settings to maximum verbatim transcript? I am able to transcribe, diarize, and get word level timestamps, but I can't figure out how to optimize for verbatim transcripts, and decrease or turn off normalization.
hotrod 1 Reputation point

2024-10-06T20:28:16.6666667+00:00
I have the same issue with my use case for verbatim transcription. I looked at the docs for Azure REST API transcription and there does not appear to be a parameter to set for turning-off/on the Azure "Defluency Removal" service ( https://learn.microsoft.com/en-us/azure/ai-services/speech-service/display-text-format?pivots=programming-language-python ).

I saw a couple of proposed methods which suggested setting the following parameter:

profanity_filter_mode = "Masked"

but that didn't work in my experience.

Wondering if I train a custom language model, would the Defluency Removal function automatically be a part of that model?

Does the Azure cognitive-services-speech-sdk ( https://github.com/Azure-Samples/cognitive-services-speech-sdk/blob/master/samples/batch/python/python-client/main.py ) provide for turning off the Defluency Removal functionality?

Share via

Transcription Denormalization.

Your answer