mms

Model type	mms
Class	Wav2Vec2ForCTCModel
Task	ctc-asr
Source	models/wav2vec2_ctc.py

Description

Wav2Vec2 CTC model for ASR — encoder + optional adapter + CTC head.

Extends Wav2Vec2Model (feature extractor + transformer encoder) with:

• Optional language-specific adapter (Wav2Vec2Adapter)

• CTC projection head (Linear → vocab logits)

Used for MMS (facebook/mms-1b-all) and any Wav2Vec2ForCTC checkpoint.

Input: input_values — (batch, num_samples) raw audio at 16 kHz Output: logits — (batch, num_frames, vocab_size) CTC logit scores

HuggingFace class: Wav2Vec2ForCTC

Language switching (MMS): Load the base model, call model.load_adapter("eng") to inject the English adapter weights, then export. The adapter + lm_head weights are language-specific; the encoder weights are shared across all languages.

Usage

mobius build --model <MODEL_ID> output_dir/

from mobius import build

model = build("<MODEL_ID>")