cstr/wav2vec2-base-960h-GGUF

wav2vec2-base-960h — GGUF

GGUF conversion of facebook/wav2vec2-base-960h for CrispASR.

This model is primarily used as a CTC forced aligner — it produces word-level timestamps by aligning a known transcript to audio. It can also perform standalone ASR via CTC greedy decode.

Architecture

• wav2vec2-base: 12-layer transformer encoder (768-dim, 12 heads)
• CTC head: Linear(768, 32) — character-level English vocabulary
• Feature extractor: 7-layer CNN (raw waveform -> 512-dim features)
• Input: 16 kHz mono audio
• Training: 960 hours of LibriSpeech (supervised CTC fine-tune)

Files

File	Size	Description
wav2vec2-base-960h.gguf	197 MB	Full precision (F16 weights, F32 norms)

Usage with CrispASR

Forced alignment (word-level timestamps)

# Align a transcript to audio — produces word-level SRT:
crispasr --align-only \
    -am wav2vec2-base-960h.gguf \
    -f audio.wav \
    --ref-text "The quick brown fox jumps over the lazy dog." \
    --align-format srt --align-output aligned.srt

# Or align from an existing SRT (re-time the cues):
crispasr --align-only \
    -am wav2vec2-base-960h.gguf \
    -f audio.wav \
    --text-file subtitles.srt \
    --align-output retimed.srt

As an aligner companion for LLM ASR backends

# Word timestamps for any backend that lacks native ones:
crispasr --backend granite -m auto \
    -am wav2vec2-base-960h.gguf --force-aligner \
    -f audio.wav -osrt

Auto-download

# The aligner is auto-downloaded when using -am auto:
crispasr --backend granite -m auto \
    -am auto --force-aligner \
    -f audio.wav -osrt

License

Apache 2.0 (same as the original model).

Credits

• Original model: Facebook AI / Meta
• GGUF conversion & runtime: CrispASR