Azure speech to text appears very slow

Question

Azure speech to text appears very slow

Sai Vishnu Soudri 65

Hi team,

We have observed that the Azure speech-to-text is very slow. I am using continuousRecognitionAsync and I observe that Azure takes a total of close to 6s for just 3s audio.

The parameters that I've set are:

EndSilenceTimeoutMs = 750

InitialSilenceTimeoutMs = 180000

The language set is en-US

I am using the standard tier and this latency feels extremely huge. Is this expected or can we make improvements to the parameters/model to get acceptable latency?

A few observations that appear as concerns:

Speech start detected takes close to 1.2s
From speech start detected, the model takes 3 more seconds to give the last recognizing response
From the last recognising response, it takes almost 2 seconds to give the recognized response.

Is continuousRecognitionAsync expected to take this much time? Please suggest some best practices for getting faster responses from Azure in real-time streaming.

Regards,

Sai Vishnu Soudri

santoshkc 15,615 Reputation points Microsoft External Staff Moderator

2025-01-30T14:56:50.2766667+00:00

Hi @Sai Vishnu Soudri,

Thank you for reaching out to Microsoft Q&A forum!

Your observed latency is expected due to processing, modelling, and network transmission. To optimize, deploy the speech resource closer to users or use Speech Containers for on-premise processing. Embedded Speech can also reduce cloud dependency.

For real-time transcription, use streaming input (Push/PullAudioInputStream) and set the default language to avoid extra processing. Asynchronous methods (start_continuous_recognition_async()) prevent blocking. Azure’s Fast Transcription (in preview) offers quicker results for large files.

For file-based transcription, split long audio into smaller chunks and process in parallel. Compressing audio before sending also reduces network delays. In speech synthesis, async methods and text streaming improve response times.

Optimizing these factors—network, streaming, async processing—can significantly reduce latency while maintaining accuracy. For more info: reduce latency.

I hope you understand. Thank you.
Anonymous

2025-01-31T14:51:01.34+00:00

Hi @Sai Vishnu Soudri,

Following up to see if the above response was helpful.
santoshkc 15,615 Reputation points Microsoft External Staff Moderator

2025-02-03T08:25:47.9533333+00:00

Hi @Sai Vishnu Soudri,

We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution, please do share that same with the community as it can be helpful to others.
scordes 5 Reputation points

2025-06-13T11:22:11.34+00:00

I have a question regarding the latency of the final result. In my case the recognizing notifications have latency (SpeechServiceResponse_RecognitionLatencyMs) of ~ 200-700ms, but the final result has a latency of ~ 2000ms.

The silence timeout is set to 200ms.

Are there any options to speed up the response of the final result?
Sven 0 Reputation points

2025-07-27T18:49:29.1666667+00:00

I experience exactly the same issues. Intermediate results appear quite fast, though.

.NET 9, Speech SDK 1.45 (latest but also reproducable with earlier versions), asynchronous/non-blocking development and continous recognition.

Appears regardless of the selected region, segmentation strategy and different timeout settings.

Compared to other STT services, the latency appears outstanding high.

I've also checked every corner of my code to rule out potential problems there. No findings and, as said, other STT solutions work pretty well.

I haven't tried to implement an Azure Speechrecognition client on my own, so I cannot say if it's a backend dependent problem or rather the SDK. However my gut feeling and general observations tell me that it could be the SDK.

Has anybody compared SDK against a custom real-time client implementation?

Your answer

santoshkc 15,615 Reputation points Microsoft External Staff Moderator

2025-01-30T14:56:50.2766667+00:00

Hi @Sai Vishnu Soudri,

Thank you for reaching out to Microsoft Q&A forum!

Your observed latency is expected due to processing, modelling, and network transmission. To optimize, deploy the speech resource closer to users or use Speech Containers for on-premise processing. Embedded Speech can also reduce cloud dependency.

For real-time transcription, use streaming input (Push/PullAudioInputStream) and set the default language to avoid extra processing. Asynchronous methods (start_continuous_recognition_async()) prevent blocking. Azure’s Fast Transcription (in preview) offers quicker results for large files.

For file-based transcription, split long audio into smaller chunks and process in parallel. Compressing audio before sending also reduces network delays. In speech synthesis, async methods and text streaming improve response times.

Optimizing these factors—network, streaming, async processing—can significantly reduce latency while maintaining accuracy. For more info: reduce latency.

I hope you understand. Thank you.
Anonymous

2025-01-31T14:51:01.34+00:00

Hi @Sai Vishnu Soudri,

Following up to see if the above response was helpful.
santoshkc 15,615 Reputation points Microsoft External Staff Moderator

2025-02-03T08:25:47.9533333+00:00

Hi @Sai Vishnu Soudri,

We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution, please do share that same with the community as it can be helpful to others.
scordes 5 Reputation points

2025-06-13T11:22:11.34+00:00

I have a question regarding the latency of the final result. In my case the recognizing notifications have latency (SpeechServiceResponse_RecognitionLatencyMs) of ~ 200-700ms, but the final result has a latency of ~ 2000ms.

The silence timeout is set to 200ms.

Are there any options to speed up the response of the final result?
Sven 0 Reputation points

2025-07-27T18:49:29.1666667+00:00

I experience exactly the same issues. Intermediate results appear quite fast, though.

.NET 9, Speech SDK 1.45 (latest but also reproducable with earlier versions), asynchronous/non-blocking development and continous recognition.

Appears regardless of the selected region, segmentation strategy and different timeout settings.

Compared to other STT services, the latency appears outstanding high.

I've also checked every corner of my code to rule out potential problems there. No findings and, as said, other STT solutions work pretty well.

I haven't tried to implement an Azure Speechrecognition client on my own, so I cannot say if it's a backend dependent problem or rather the SDK. However my gut feeling and general observations tell me that it could be the SDK.

Has anybody compared SDK against a custom real-time client implementation?