Changelog - V5 just released!

[snakers4]

Just a handy issue to be notified of latest changes and micro-releases (we will mostly changing the models)

[snakers4]

Initial models, examples, utils for VAD only uploaded (no number detector or language classifier yet)

[snakers4]

First readable public release

[snakers4]

Added VAD latency and throughput metrics

[snakers4]

Updated VAD quality Before / after (precision / recall)

[adamnsandle]

Added < 250ms compatibility

[Sontref]

Added number detector

[snakers4]

Language detector example, readme update + FAQ

[snakers4]

Audiotok benchmarks added Looks like all energy based solutions are kind of similar

[snakers4]

Added a utility to tune the VAD params properly for a domain

[snakers4]

Some final benchmarks posted here - https://github.com/pyannote/pyannote-audio/issues/604#issue-798003383 Probably we are done with benchmarks for now

[snakers4]

Added micro (10k params, 100x smaller) VAD models

[snakers4]

Added micro (10k params, 100x smaller) VAD models for 8 kHz audio

[snakers4]

• Added mini (100k params) VAD models for 8 kHz and 16 kHz
• Added adaptive vad iterator

https://github.com/snakers4/silero-vad/pull/54

[snakers4]

• Fixed examples and notebooks
• Updated README
• Added adaptive examples

[snakers4]

• Added a language classifier for 116 languages
• It classifies audios into languages and mutually intelligible language groups (i.e. Serbian + Bosnian + Croatian, Russian + Ukranian + others, Hindi + Urdu, etc), see the full list here and here
• Probably some artifical / unspoken languages will be excluded and a large model will be trained

[snakers4]

improved language classifier

• 95 languages (85% accuracy), 58 language groups (90% accuracy)
• Mutually intelligible languages are united into language groups (i.e. Serbian + Croatian + Bosnian are very similar)
• Trained on approx 20k hours of data (10k of which are for 5 most popular languages)
• 4.7M params

[snakers4]

updated further reading section

[snakers4]

New V3 Silero VAD is Already Here

Main changes

• One VAD to rule them all! New model includes the functionality of the previous ones with improved quality and speed!
• Flexible sampling rate, 8000 Hz and 16000 Hz are supported;
• Flexible chunk size, minimum chunk size is just 30 milliseconds!
• 100k parameters;
• GPU and batching are supported;
• Radically simplified examples;

Migration

Please see the new examples.

New get_speech_timestamps is a simplified and unified version of the old deprecated get_speech_ts or get_speech_ts_adaptive methods.

speech_timestamps = get_speech_timestamps(wav, model, sampling_rate=16000)

New VADIterator class serves as an example for streaming tasks instead of old deprecated VADiterator and VADiteratorAdaptive.

vad_iterator = VADIterator(model)
window_size_samples = 1536

for i in range(0, len(wav), window_size_samples):
   speech_dict = vad_iterator(wav[i: i+ window_size_samples], return_seconds=True)
   if speech_dict:
       print(speech_dict, end=' ')
vad_iterator.reset_states()

[snakers4]

Even Better V3 Silero VAD

• Models with even higher quality (just see the plots with metrics!);
• New model ~ large model >> all previous (even large) models;
• Now model works properly quality-wise, i.e. 100ms > 60ms > 30ms and16 kHz > 8 kHz;

[snakers4]

This summarises new progress well

[snakers4]

New V3 ONNX VAD Released

We finally were able to port a model to ONNX:

• Compact model (~100k params);
• Both PyTorch and ONNX models are not quantized;
• Same quality model as the latest best PyTorch release;
• Only 16kHz available now (ONNX has some issues with if-statements and / or tracing vs scripting) with cryptic errors;
• In our tests, on short audios (chunks) ONNX is 2-3x faster than PyTorch (this is mitigated with larger batches or long audios);
• Audio examples and non-core models moved out of the repo to save space;

[snakers4]

Support For Sampling Rates Higher Than 16 kHz

• jit model now can handle 8, 16, 32 and 48 kHz directly (change implemented within the model itself);
• onnx model as well, but only via external wrappers (we just use each n-th sample for higher sampling rates);
• This support is mostly a hack, i.e. we just use each n-th sample for higher sampling rates (instead of averaging);

[snakers4]

⚠️ Important Information for VAD Python Users ⚠️

If you are using the VAD in a:

• multi-threaded or
• a multi-process application

Do not forget to disable gradients in EACH process and / or thread. Otherwise memory may leak noticeably.

[snakers4]

[adamnsandle]

New V4 VAD Released

Changes:

• Improved quality
• Improved perfomance
• Both 8k and 16k sampling rates are now supported by the ONNX model
• Batching is now supported by the ONNX model
• Added audio_forward method for one-line processing of a single or multiple audio without postprocessing

[snakers4]

It is worth posting this chart:

[snakers4]

• Remove picovoice mentions

[snakers4]

• Deprecate language classifier and number detector models, since they are not maintained anymore.