Property	Value
Architecture	Wav2Vec2 (CNN feature extractor + 24 Transformer layers)
Hidden size	1024
Attention heads	16
CTC vocabulary	33 tokens
Format	GGUF (F16 weights)
Size	627 MB

The model was pre-trained on 53 languages (XLSR-53) and fine-tuned on English Common Voice data with a CTC head. It accepts 16 kHz mono audio and outputs character-level transcriptions.

Usage with CrispASR

bash

crispasr \
  --backend wav2vec2 \
  -m wav2vec2-xlsr-en.gguf \
  audio.wav

Provenance

Weights were converted from the original Hugging Face PyTorch checkpoint into GGUF format with F16 precision for all transformer and feature-extractor parameters.

License

Apache-2.0 — same as the original model.

cstr/wav2vec2-large-xlsr-53-english-GGUF

Wav2Vec2 Large XLSR-53 English — GGUF

Model details

Usage with CrispASR

Provenance

License

cstr/
wav2vec2-large-xlsr-53-english-GGUF