Is it possible to change speech recognition parameters in "Recognizing" or "Recognized" handlers?

Faris Lemes 20

Hi I'm having the callbacks for Recognizing and Recognized handlers for the speech recognition, also, I have keyword recognition and continues recognition. Is there a possibility to update recognition parameters in those callbacks? Use case scenario is for example when keyword is detected I want to be able to make some longer pause (3-4 seconds) after keyword then continue to speak and recognizer should recognize it. After I start talking (after keyword) I don't want to have the same pause after my last word before recognition is over. I was trying to achieve this by stopping the recognition inside Recognized handler, updating the parameter PropertyId::Speech_SegmentationSilenceTimeoutMs and resuming the recognition when I get response in handler:

ResultReason::RecognizedKeyword

I've tried to use this for updating parameters https://learn.microsoft.com/en-us/answers/questions/1643265/how-to-change-speech-recognition-configuration-aft but I assume that there is some better way of achieving this. Is there any integrated way in speech SDK to achieve this?

Thanks!

YutongTie-MSFT 46,986 Reputation points

2024-04-09T21:24:12.5033333+00:00

@Faris Lemes Thanks for reaching out to us.

As of the current version of the Azure Speech SDK, there is no built-in way to change the speech recognition parameters such as the silence timeout during the Recognizing or Recognized event handlers.

The parameters are typically set at the start when you create a speech recognizer instance and cannot be modified once the recognizer has started processing audio. Your approach of stopping the recognizer, changing the parameters, and then restarting the recognizer could work, but it may introduce some complexity and potential delay in the recognition process, especially if you're working with continuous recognition.

One possible workaround could be to use two separate speech recognizers: one for keyword detection with a longer silence timeout, and another for the actual speech recognition with a shorter silence timeout. When the keyword is detected, you could stop the keyword recognizer and start the speech recognizer.

Please let me know if this workaround works for you.

Regards,

Yutong
Faris Lemes 20 Reputation points

2024-04-12T06:47:12.9266667+00:00

I've tried to update parameters from the additional thread. In callbacks I would set some flags which were indication for this separate thread to update the parameters in specific way. The behavior in this case is that when I say sentence (longer one) for example, speech recognizer would update the timeout and after I stop speaking silence will be shorter. But there is one issue, for example if I have PropertyId::Speech_SegmentationSilenceTimeoutMs set to 4000 for example, and if I detect for example yes or no, or some short text and there is silence afterwards, and I want to have this silence as small as possible (500 ms) it would not be changed. I have a feeling that recognizer behaves different for longer and for shorter sentences. Are you familiar with this behavior?