Silero VAD v6 as safetensors

Silero VAD 6.2.3, 16 kHz branch, as safetensors for scriptrs, a Rust speech-to-text library. The weights are unchanged.

scriptrs runs it natively in Rust on the CPU, at about 0.5 s per hour of audio, to plan long-form transcription around speech regions when its cuda and vad features are enabled. On a GPU, each step of the model's recurrence would cost more in kernel launches than the whole step costs on the CPU.

By default scriptrs runs it together with Silero VAD v5 and keeps each window's higher speech probability, at a 0.6 threshold. Either model alone misses speech the other finds: v6 separates speech from non-speech better on most audio, and v5 holds up better on noisy dinner-party audio.

Accuracy

Parakeet TDT 0.6B v3 on an RTX 5090, long-form VAD planning, WER by VAD model:

v5 and v6 alone use a 0.5 speech threshold. The sparse set is LibriSpeech test-clean utterances separated by 15 to 75 s of film music, effects, ambience and low-level noise.

Usage

scriptrs = { version = "0.5", features = ["cuda", "vad"] }

use scriptrs::{
    LongFormConfig, LongFormMode, LongFormTranscriptionPipeline, ModelBundle, ParakeetDeployment,
};

// downloads Parakeet and both Silero VAD repositories at pinned revisions on first use
let bundle = ModelBundle::from_pretrained_native_long_form(ParakeetDeployment::v3())?;
let pipeline = LongFormTranscriptionPipeline::from_bundle(bundle)?;
let config = LongFormConfig {
    mode: LongFormMode::Vad,
    ..LongFormConfig::default()
};
let result = pipeline.run_with_config(&audio, &config)?; // mono 16 kHz f32 samples

To run v6 alone, also set segmentation: VadSegmentationConfig::for_model(VadModel::SileroV6). scriptrs 0.5.0 and later download from this repository.

Files

The file records its source, the source model's SHA-256 (e1122837f4154c511485fe0b9c64455f7b929c96fbb8d79fbdb336383ebd3720) and its config in its metadata.

License

Silero VAD is licensed under the MIT License, copyright Silero Team. See LICENSE.