Finally, V5 is here, 3x faster, supporting 6000+ languages!

Performance and Model Size

• 3x faster inference for TorchScript, 10% faster inference for ONNX;
• Now TorchScript is as fast as ONNX;
• Model size is 2x larger, 2MB vs. 1MB;

Quality

• The VAD supports more than 6,000 languages now;
• Significanly more robust on noisy data;
• Overall 5-7% quality increase on clean data;
• Quality difference for 8 kHz and 16 kHz is negligible now;
• Quality difference for different window sizes is negligible => window size was deprecated;
• Added benchmarks on 9 unique datasets (2 private) and one holistic multi-domain dataset;

Changes and deprecations

• ONNX opset 16;
• window_size_samples is deprecated - now the VAD only works with fixed size window;
• VAD now works with 8 kHz and 16 kHz sample rates, only with fixed 256 and 512 sample windows respectively;
• Slightly changed internal logic, now some context (part of previous chunk) is passed along with the current chunk;
• Sample rates that are a multiple of 16 kHz are still supported;