Is it possible to recognize short words ("Yes", "No", "Ok") in speech sdk consistently

Faris Lemes 50

hello, I was experimenting with SPEECH SDK and I was using https://github.com/Azure-Samples/cognitive-services-speech-sdk/blob/master/quickstart/cpp/linux/from-microphone/helloworld.cpp

I've adjusted code a bit, I was using the following timeouts:

config->SetProperty(PropertyId::SpeechServiceConnection_InitialSilenceTimeoutMs, "1500");
config->SetProperty(PropertyId::Speech_SegmentationSilenceTimeoutMs, "5000");
config->SetProperty(PropertyId::SpeechServiceConnection_EndSilenceTimeoutMs, "1000");
auto audioProcessingOptions = AudioProcessingOptions::Create(AUDIO_INPUT_PROCESSING_ENABLE_DEFAULT);
auto recognizer = SpeechRecognizer::FromConfig(config, audioInput);

One thing that I noticed is that when I say only some short words i.e. "yes", "no", "ok" I would get weird results from the picture image.png. I was not saying anything else except "Yes", "No" and "OK".

Thanks

dupammi 7,955 Reputation points Microsoft Vendor

2024-06-03T07:05:12.0366667+00:00
Hi @Faris Lemes

Thank you for your question.

Yes, it is possible to recognize short words like "Yes", "No", and "Ok" consistently using the Speech SDK. One thing you can try is to add a phrase list with common short commands to your recognizer setup. Here's an example of how you might adjust your C++ code to include a phrase list:

auto config = SpeechConfig::FromSubscription("YourSubscriptionKey", "YourServiceRegion"); // Add short words to a phrase list auto phraseList = PhraseListGrammar::FromRecognizer(recognizer); phraseList->AddPhrase("Yes"); phraseList->AddPhrase("No"); phraseList->AddPhrase("Ok"); // Adjust timeout values as needed config->SetProperty(PropertyId::SpeechServiceConnection_InitialSilenceTimeoutMs, "1500"); config->SetProperty(PropertyId::Speech_SegmentationSilenceTimeoutMs, "5000"); config->SetProperty(PropertyId::SpeechServiceConnection_EndSilenceTimeoutMs, "1000"); auto recognizer = SpeechRecognizer::FromConfig(config, audioInput); cout << "Say something...\n"; auto result = recognizer->RecognizeOnceAsync().get(); // Handle the result as before...

This should help improve the recognition accuracy for the specified short words. If your use case involves multiple short utterances in quick succession, you might also consider using continuous recognition instead of RecognizeOnceAsync(). And be sure to adjust the timeout values as needed to handle short pauses between commands but not cut off recognition prematurely.

Please refer : https://github.com/Azure-Samples/cognitive-services-speech-sdk/blob/master/samples/csharp/sharedcontent/console/speech_recognition_samples.cs

Let me know if this helps. Thank you.
Faris Lemes 50 Reputation points

2024-06-03T07:09:14.38+00:00

I've already tried to add phraseList, I will try continuous recognition as well, thanks for the advice

Accepted answer

Gowtham CP 3,730 Reputation points

2024-06-03T05:21:18.5666667+00:00
Hello Faris Lemes ,

Thanks for reaching out in the Microsoft Q&A!

Here are some tips to improve recognition:

Increase Timeout: Try setting the "Speech_SegmentationSilenceTimeoutMs" to 3 seconds to give the system more time to process short words.

Clear Speech: Ensure you're in a quiet environment and enunciate clearly to avoid confusion.

Update SDK: Make sure you’re using the latest version of the Speech SDK to benefit from recent fixes.

Custom Model: If these words are super important, consider training a custom speech model with your specific vocabulary. It'll learn your voice and these keywords even better.

Please let me know if you have further questions.

If the response helped, do "Accept Answer" and up-vote it to close the thread! Thanks
Please sign in to rate this answer.
Faris Lemes 50 Reputation points

2024-06-03T05:31:11.5366667+00:00

If I increase the timeout, speech SDK will wait for 3 seconds after my last word to detect that speaking was finished. Which is not acceptable in my case, I'm trying to make some voice assistant. I had this timeout higher before, and duration was not acceptable.

Faris Lemes 50 Reputation points

2024-06-03T05:43:23.43+00:00

For steps 2 and 3 I'm using latest version of the Speech SDK and I was in quiet room.
Only thing which is left is probably custom model then?

Gowtham CP 3,730 Reputation points

2024-06-03T06:46:08.86+00:00

Hello Faris Lemes ,

As far as I know, there aren’t any hidden settings beyond the steps you've tried.

Since you’re already using the latest SDK and speaking clearly in a quiet environment, a custom speech model is the most promising approach. Training a custom model specifically for short commands like "yes," "no," and "ok" will significantly improve recognition accuracy for your voice assistant. Thanks!

Faris Lemes 50 Reputation points

2024-06-03T06:47:54.48+00:00

thanks for the answers
Sign in to comment

Share via

Is it possible to recognize short words ("Yes", "No", "Ok") in speech sdk consistently

0 additional answers