Include custom audio files for keyword recognition training process

Question

I am leveraging Azure Keyword Recognition service, it works pretty nice except some false wakeup.

We've collected a bunch of false waking up audio files, and I was wondering whether there is some approach that we can include these false audio files into the training process.

Any advice will be appreciated!

Thanks in advance!

Answer 1

Hi @Liu Wenbin (Lofty Team)

Thank you for using the Microsoft Q&A forum.

To address the false wakeup issue in Azure Keyword Recognition service, you can include your collected false waking up audio files into the training process. To do this, you need to prepare your custom audio files, create human-labeled transcriptions, upload your audio files and transcriptions to the training dataset, train your model using Speech Studio, and ensure that your audio files and transcriptions are in one of the supported languages (English (United States) and Chinese (Mandarin, Simplified)). By following these steps, you can integrate your custom audio files into the created custom keyword recognition training process, potentially reducing false wakeup instances.

That said, to create a custom keyword model, the text files or transcripts play a crucial role, so it's important to ensure that the correct text is added. The audio files complement the text files by helping train the model based on your audio quality or background, which can be used with all your future files. It's preferable to use short audio files for training the acoustic model. For more details, refer - Audio + human-labeled transcript data for training or testing.

I hope the provided information helps! Thank you.

---

Please don't forget to click Accept Answer and Yes for was this answer helpful.