cstr/wav2vec2-xls-r-300m-cs-250-GGUF

Wav2Vec2 Xls R 300M Cs 250 -- GGUF

GGUF conversions and quantisations of comodoro/wav2vec2-xls-r-300m-cs-250 for use with CrispStrobe/CrispASR.

Available variants

File	Quant	Size	Notes
wav2vec2-xls-r-300m-cs-250-f16.gguf	F16	~627 MB	Reference conversion
wav2vec2-xls-r-300m-cs-250-q4_k.gguf	Q4_K	~212 MB	Default auto-download target
wav2vec2-xls-r-300m-cs-250-q8_0.gguf	Q8_0	~356 MB	Higher precision

Model details

• Architecture: Wav2Vec2ForCTC
• Hidden size: 1024
• Attention heads: 16
• Transformer layers: 24
• CTC vocabulary: 46 tokens
• Language: Czech
• Base model: comodoro/wav2vec2-xls-r-300m-cs-250

Usage with CrispASR

crispasr --backend wav2vec2 -m wav2vec2-xls-r-300m-cs-250-q4_k.gguf -f audio.wav -l cs

# Auto-download via registry alias once published:
crispasr --backend wav2vec2 -m auto --auto-download -l cs -f audio.wav

Provenance

Converted from the upstream Hugging Face checkpoint with models/convert-wav2vec2-to-gguf.py, then quantized with build-ninja-compile/bin/crispasr-quantize.

License

Apache-2.0 — same as the upstream model card.