cstr/reazonspeech-nemo-v2-GGUF

ReazonSpeech NeMo v2 (Japanese) — GGUF

GGUF / ggml conversions of reazon-research/reazonspeech-nemo-v2 for use with the crispasr CLI from CrispStrobe/CrispASR.

A 619 M-parameter Japanese ASR model trained on the ReazonSpeech v2.0 corpus (~35,000 hours of Japanese audio):

• FastConformer-RNNT — pure RNN-Transducer decoder (no TDT duration head).
• Local relative-position attention (window 128 + 128, plus 1 global token), so the encoder scales to long audio without quadratic blow-up.
• 80-mel front-end, 16 kHz mono, 3000-token SentencePiece vocabulary.
• Apache-2.0 licence.

Files

File	Size	Notes
reazonspeech-nemo-v2-f16.gguf	1.24 GB	F16 — highest fidelity, closest to the NeMo reference
reazonspeech-nemo-v2-q8_0.gguf	738 MB	Q8_0 — default download, near-F16 quality
reazonspeech-nemo-v2-q4_k.gguf	477 MB	Q4_K — smallest; some accuracy loss, fine for quick checks

Q8_0 is the recommended general-purpose quant; use F16 when you want the closest match to the official NeMo Python pipeline.

Quick start

# 1. Build the runtime
git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc) --target crispasr

# 2. Download the Q8_0 (default) — or swap the filename for the F16 / Q4_K
huggingface-cli download cstr/reazonspeech-nemo-v2-GGUF \
    reazonspeech-nemo-v2-q8_0.gguf --local-dir .

# 3. Transcribe a 16 kHz mono WAV
./build/bin/crispasr --backend parakeet \
    -m reazonspeech-nemo-v2-q8_0.gguf -f your-japanese-audio.wav -t 8

crispasr can also fetch the model for you by its registry name:

./build/bin/crispasr --backend parakeet -m reazonspeech \
    --auto-download -f your-japanese-audio.wav

(Both this RNNT model and the sibling cstr/parakeet-ctc-1.1b-ja-GGUF run through crispasr's parakeet backend — the runtime selects the RNNT vs. CTC decode path from the GGUF metadata.)

Long-form audio

The local-attention encoder handles long inputs, but as with the other Japanese FastConformer models a single long pass can drift; for clips longer than ~15 s prefer VAD-bounded chunking:

./build/bin/crispasr --backend parakeet -m reazonspeech-nemo-v2-q8_0.gguf \
    -f long-japanese-audio.wav --vad -t 8

Model architecture

Component	Details
Encoder	FastConformer with local relative-position attention (window 128+128, 1 global token)
Decoder	RNN-Transducer (RNNT) — LSTM predictor + joint network; no TDT durations
Vocab	3000 SentencePiece tokens (Japanese)
Audio	16 kHz mono, 80 mel bins, n_fft=512, hop=160, win=400
Parameters	~619 M

How this was made

1. The .nemo checkpoint from reazon-research/reazonspeech-nemo-v2 is unpacked; architecture hyperparameters (d_model, layers, local-attn window, predictor/joint dims, vocab) are read from model_config.yaml and cross-checked against the tensor shapes. The mel filterbank and Hann window are baked into the GGUF so the runtime reproduces NeMo's front-end.
2. NeMo state-dict keys are remapped to ggml-friendly names — matmul tensors as F16, norms / biases / mel filterbank as F32 — and the F16 GGUF is quantised to Q8_0 and Q4_K.
3. Inference runs through src/parakeet.{h,cpp} in CrispASR, which handles the local relative-position attention and the RNNT predictor/joint loop.

Licence

Apache-2.0, inherited from reazon-research/reazonspeech-nemo-v2. Please also see the ReazonSpeech project for details on the training corpus.