Loading page
Checkpoint options and configuration for Wav2Vec2 models.
AudioLens's Wav2Vec2 model is used for emotion classification, not transcription.
| Catalog key | Model ID | Task | Notes |
|---|---|---|---|
wav2vec2 | r-f/wav2vec-english-speech-emotion-recognition | emotion classification | shown in the UI dropdown; labels are read from the model's config at runtime |
Audio is resampled to 16 kHz before inference. For each clip the model
returns a probability for every emotion class, the top class as the
predicted emotion, and its probability as the confidence. The classes
are read from the model's config at runtime; for the bundled checkpoint they
are neutral, happy, sad, angry, and fear.
Like Whisper, this model is English-only — there is no multi-language support for emotion classification.
Two older Wav2Vec2 checkpoints remain in the codebase as a legacy fallback path, used only for attention visualization on an older code path:
jonatasgrosman/wav2vec2-large-xlsr-53-englishfacebook/wav2vec2-base-960hBoth are Apache 2.0 licensed. They aren't part of the main model catalog and aren't offered in the UI dropdown.