Hindi Wav2Vec2 Aligner — q8_0 GGUF for CrispASR

Quantized GGUF of theainerd/Wav2Vec2-large-xlsr-hindi (wav2vec2-large XLSR-53 fine-tuned on Hindi, 315M parameters) for use as the CTC word-timestamp aligner in CrispASR.

It is meant to be paired with the Hindi checkpoints of Bonenk/indic-transcribe-flex-GGUF. That repo's built-in -am auto aligner (canary-ctc-aligner) is English-oriented, so its word boundaries drift on Hindi transcription. This checkpoint replaces it with a Devanagari CTC model, giving accurate per-word timestamps for Hindi output.

This is a community conversion, not an official release.

Files

File	Size	Notes
`hi-wav2vec2-q8_0.gguf`	356 MiB	Q8_0 aligner GGUF; tokenizer embedded, no separate download needed

Usage

Build CrispASR from the indic-transcribe-flex branch (see the main repo README), then download both files:

hf download Bonenk/indic-transcribe-flex-GGUF indic-transcribe-flex-q5_k.gguf --local-dir .
hf download Bonenk/hi-wav2vec2-aligner-GGUF hi-wav2vec2-q8_0.gguf --local-dir .

Pass the aligner explicitly with -am and keep --force-aligner so its timestamps are used even though the Canary backend emits its own:

./build/bin/crispasr --backend canary \
  -m indic-transcribe-flex-q5_k.gguf \
  -f hindi.wav -l hi \
  -am hi-wav2vec2-q8_0.gguf --force-aligner \
  --output-srt

Tip: on Windows, refer to the binaries in the same folder (for example, from this repo's build):

crispasr.exe --backend canary ^
  -m indic-transcribe-flex-q5_k.gguf ^
  -f hindi.wav -l hi ^
  -am hi-wav2vec2-q8_0.gguf --force-aligner ^
  --output-srt

Notes:

-l hi is still required by the Canary backend; the aligner is only used for word timestamps.
Drop --force-aligner to compare native Canary timestamps against the CTC aligner's.
Works with other LLM-style CrispASR backends (--backend qwen3, --backend cohere, ...) where -am is otherwise the only way to get word timestamps; the aligner needs the transcript's script to match its Devanagari vocabulary.
The aligner runs standalone, so the two models can run on different backends (e.g. --gpu-backend cuda for the Canary model and --no-gpu for the aligner).

Model architecture

Key	Value
Architecture	`wav2vec2` (GGUF v3, 422 tensors)
Parameters	~315M (Q8_0 + F32 CTC head/norms)
Hidden size	1024
Layers / heads	24 / 16
Feature extractor	7 conv layers, 512 channels
Vocab	64-token Devanagari CTC charset
Context stride	5 conv + 5x2^4 strides → ~20 ms per frame

Conversion

Converted with CrispASR's models/convert-wav2vec2-to-gguf.py from a full-precision checkout of the source checkpoint, then quantized with crispasr-quantize q8_0:

python models/convert-wav2vec2-to-gguf.py \
  --model-dir theainerd/Wav2Vec2-large-xlsr-hindi \
  --output hi-wav2vec2-f16.gguf

./build/bin/crispasr-quantize hi-wav2vec2-f16.gguf hi-wav2vec2-q8_0.gguf q8_0

The quantized tokenizer in this repo reproduces the source model's 64-token vocab.json exactly (verified token-for-token).

Validation

Informal checks with CrispASR:

Word timestamps from this aligner on Hindi audio are visibly tighter than the default canary-ctc-aligner path, especially for words transcribed as short Devanagari compounds.
No WER benchmark was produced — this checkpoint is intended for alignment only, not as an ASR backend.

License and attribution

The base model theainerd/Wav2Vec2-large-xlsr-hindi is inherited from facebook/wav2vec2-large-xlsr-53 (Apache-2.0); the converted weights are distributed under the same terms.

Source model: theainerd/Wav2Vec2-large-xlsr-hindi
Base model: facebook/wav2vec2-large-xlsr-53
Companion ASR models: Bonenk/indic-transcribe-flex-GGUF
Runtime: CrispASR (MIT)