planning: Ichigo VAD

[dan-homebrew]

Goal

• Remove the need to press the button, detect the voice
  • Medium-term
  • Enables ambient voice detection
  • Enables interruptibility
• Small model that has binary classifier for voice activity detection
  • Need to determine activity parameter (e.g. how many seconds)
• Long-term: advance VAD to a listening model

Tasklist

• [ ] Menlo Realtime API
• [ ] Cortex Realtime API

Resources

• Use SileroVAD https://github.com/snakers4/silero-vad used by both VITA and Huggingface speech2speech model
• Huggingface pipeline might be a good reference point https://github.com/huggingface/speech-to-speech

[hahuyhoang411]

e2e vad: https://github.com/modelscope/FunASR

[PodsAreAllYouNeed]

FunASR is used by huggingface to support the Paraformer STT model, while they use SileroVAD. The FSMN-VAD provided by FunASR could be useful to look into as well. Also the pipeline for FunASR includes VAD and Diarization together with STT which could indeed be very useful.

The VAD handler written by hf using some of the SileroVAD code is quite nice: https://github.com/huggingface/speech-to-speech/blob/93d74ba3bc3ad1a948cc167d7cdb95699e49d867/VAD/vad_handler.py

It includes enhancement as well, which is very useful. We can potentially adapt the handler to support other VADs as well. This can cater to #93 as well.

Current Pipeline Audio -> Ichigo -> TTS

Pipeline using hf/s2s handler Audio -> (VAD -> Enhancement) -> Ichigo -> TTS