Starstruk Logo MarkStarstruk

v5.0

snakers4/silero-vadv5.0Jun 27, 2024by snakers4

AI Summary

Version 5.0 delivers a massive performance boost with 3x faster inference and support for over 6,000 languages, while significantly improving robustness on noisy data and offering a 5-7% quality increase on clean audio.

Key Highlights

  • 3x faster inference for TorchScript and 10% faster for ONNX, with TorchScript now matching ONNX speeds.
  • Support for over 6,000 languages.
  • Added benchmarks on 9 unique datasets (2 private) and one holistic multi-domain dataset.
  • Quality improvement of 5-7% on clean data with significantly better robustness on noisy data.
  • Model size increased to 2MB (from 1MB).

Breaking Changes

  • `window_size_samples` parameter is deprecated and removed, forcing use of fixed window sizes (256 or 512 samples).
  • VAD now strictly supports 8 kHz and 16 kHz sample rates (or multiples of 16 kHz) with specific fixed window sizes.
  • Internal logic changed to pass context from previous chunks.

New Features

  • Support for over 6,000 languages
  • Added benchmarks on 9 unique datasets (2 private) and one holistic multi-domain dataset
  • Increased robustness on noisy data
  • Quality improvement of 5-7% on clean data

Full Release Notes

![image](https://github.com/snakers4/silero-vad/assets/12515440/5f581592-3035-411e-9803-44a2fda0af8e)

## **Performance and Model Size**

- 3x [faster](https://github.com/snakers4/silero-vad/wiki/Performance-Metrics) 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 ](https://github.com/snakers4/silero-vad/wiki/Quality-Metrics) 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;