Loading page
Checkpoint options and configuration for Whisper models.
AudioLens ships two Whisper checkpoints for speech-to-text, both under the Apache 2.0 license.
| Catalog key | Model ID | Notes |
|---|---|---|
whisper-base | openai/whisper-base | shown in the UI model dropdown |
whisper-large | openai/whisper-large-v3 | API-only — intentionally hidden from the UI dropdown |
whisper-large is fully usable through the REST API (POST /jobs, POST /models, etc.) — it's just not offered as a choice in the frontend's model
picker.
Whisper decoding in AudioLens is pinned to language="en" and
task="transcribe". AudioLens does not support multi-language transcription
or translation tasks — every Whisper job runs in English-transcribe mode
regardless of the input clip's actual language.
Whisper models can be analyzed with the Saliency panel. One thing worth
knowing before you rely on the result: the method labeled gradcam in the
API and UI does not actually run Grad-CAM — it runs Integrated Gradients
under the hood. See /docs/saliency for the full detail on
saliency methods and how to check whether a given map is trustworthy.