QUESTION: I am looking for a debugging tool/method that I can use to discover what the problem is when PocketSphinx.js does not recognize a custom word I have defined. Specifically, I would like to be able to see what phonemes that PocketSphinx.js is actually hearing when a word is not recognized.
Details:
I am working on developing a voice interface for a website. I need to use a custom keyword, and I am working with PocketSphinx.js. The custom keyword I've chosen is "Berk-o-bot" with a CMU pronunciation of "B ER K OW B AA T". I've had very low success getting this phrase to recognize. (Recognizes only about a quarter of the time I slowly and deliberately say the word, never if I speak at normal speed.) NOTE:
If I split the word into smaller chunks (Berk [B ER K], Berk-O [B ER K OW], O-bot [OW B AA T] or bot [B AA T]), the smaller chunks recognize easily.
Thoughts from failed attempts:
- My suspicion is that the recognizer has difficulty since the first and thirds syllables are accented, but not the 2nd. I have tried using adding Arpabet Stress Symbols to the pronunciation [B ER1 K OW0 B AA2 T] but pocketsphinx.js threw errors when I would try to add the word to the dictionary.
- I've also tried adding multiple pronunciations to the dictionary like BERK-O-BOT(2) [B ER K AO B AA T], BERK-O-BOT(3) [B ER K AH B AA T], BERK-O-BOT(4) [B ER K EH B AA T], BERK-O-BOT(5) [B ER K UH B AA T] but did not see any improvement in recognition.
- I am new to pocketsphinx.js, so the above two methods might work, but I might be implementing them incorrectly.
You need to make it working on desktop with pocketsphinx_continuous and pre-recorded audio files first. Once it works you can try in a web browser with the same configuration and models.
You can share audio files and the configuration/models if it still doesn't work to get more detailed help.