dwks / silvius

Kaldi-based speech recognition system + grammar
http://voxhub.io/silvius
BSD 2-Clause "Simplified" License
100 stars 28 forks source link

Improving recognition accuracy #32

Open lpww opened 5 years ago

lpww commented 5 years ago

Hey, thanks for making this amazing tool! I think it could work well for me but I'm running into some issues with the speech recognition and I'm hoping to get some input on the best resolution.

Some of my commands are recognized first time but most require multiple repeats and some are never recognized regardless of how many times I repeat. I've tried both of the public services and the beta is definitely better but still not usable.

I think the issue could be my English accent or my microphone quality. I'm using a hyperx cloud silver gaming headset that I assumed would have a decent enough mic but maybe not. What do you think?

These are the mic specs:

* Element: Electret condenser microphone
* Polar pattern: Uni-directional, noise-cancelling
* Frequency response: 50Hz-18,000 Hz
* Sensitivity: -39dBV (0dB=1V/Pa,1kHz)
dwks commented 5 years ago

Sorry for the delay. I'm not familiar with that microphone, it looks decent. The speech model is probably having trouble with your voice, it may be your accent, or other properties of your speech. In order to get a working speech system for you, can you send me the complete list of all words that work well and ones that don't? I will try to see if there is some sort of pattern which would indicate a certain phone is not being recognized. In the worst case, I can train a new model for you using only words that the system recognizes well in your voice.

lpww commented 5 years ago

Wow, that would be amazing, thanks! I'll do some more thorough testing and get you a full list. I forgot to say it but I also sound very congested due to chronic allergies, which could be another reason for being difficult to understand.