Custom checkpoints

Bring your own model

Architecture constraints

A custom model must be one of:

• a speech Seq2Seq architecture (like Whisper),
• a CTC architecture, or
• an audio-classification architecture (like Wav2Vec2 for emotion).

Two rules apply to every custom model, no exceptions:

• trust_remote_code=False always — a repo that requires remote code execution to load cannot be registered.
• Classification models must define id2label in their config, so AudioLens knows what the output classes mean. A classification checkpoint without id2label will fail validation.

Registering a model

Registering a custom model follows the same asynchronous pattern as running an analysis job:

1. 1
   Submit the model

   POST /models with the Hugging Face repo id. This returns immediately with a 202 while the model is validated in the background.

2. 2
   Validation runs

   AudioLens checks the model's architecture and config against the constraints above — Seq2Seq, CTC, or audio-classification; trust_remote_code=False; id2label present for classification models.

3. 3
   Poll for the result

   GET /models/{id} to check status. Once validation succeeds, the model is available to use in jobs like any built-in model.

Once registered, list, inspect, or remove custom models with GET /models, GET /models/{id}, and DELETE /models/{id}. If the model supports it, GET /models/jacobian-lenses/{model_id} lists any Jacobian Lenses fitted for it.

Where to go next