Ridiculously fast, only slightly bad voice activity detection in pure Rust. Port of the famous WebRTC VAD.
#![no_std]
, doesn't even requirealloc
- Internal buffers can get pretty big when stored on the stack, so the
alloc
feature is enabled by default, which allocates them on the heap instead.
- Internal buffers can get pretty big when stored on the stack, so the
- Stupidly fast; uses only fixed-point arithmetic
- Achieves an RTF of ~3e-4 with 30 ms 48 KHz frames, ~3e-5 with 30 ms 8 KHz frames.
- Comparatively, Silero VAD v4 w/
ort
achieves an RTF of ~3e-3 with 60 ms 16 KHz frames.
- Okay accuracy
- Great at distinguishing between silence and noise, but not between noise and speech.
- Earshot provides alternative models with slight accuracy gains compared to the base WebRTC model.