cstr/parakeet-ctc-1.1b-ja-GGUF

Parakeet CTC 1.1B (Japanese) — GGUF

GGUF / ggml conversions of `grider-transwithai/parakeet-ctc-1.1b-ja` for use with the crispasr CLI from [CrispStrobe/CrispASR](https://github.com/CrispStrobe/CrispASR).

A 1.1 B-parameter Japanese ASR model:

• —FastConformer-CTC — a 42-layer FastConformer encoder with a CTC decoder (greedy CTC at inference; one linear head over the SentencePiece vocabulary, no RNNT/TDT predictor).
• —Fine-tuned from NVIDIA's English `nvidia/parakeet-ctc-1.1b` on Japanese data.
• —80-mel front-end, 16 kHz mono, 8× temporal subsampling (50 → 12.5 fps).
• —Apache-2.0 licence (the NVIDIA base architecture is CC-BY-4.0).

Files

For a CTC model the Q80 quant is robust (CTC is far less sensitive to quantisation noise than the small JA TDT decoder, which can loop). Use **Q80 for general transcription and F16** when you want the closest match to the 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/parakeet-ctc-1.1b-ja-GGUF \
    parakeet-ctc-1.1b-ja-q8_0.gguf --local-dir .

# 3. Transcribe a 16 kHz mono WAV
./build/bin/crispasr \
    -m parakeet-ctc-1.1b-ja-q8_0.gguf -f your-japanese-audio.wav -t 8

Backend: this is a CTC model — let crispasr auto-detect it (as above, no --backend) or pass --backend fastconformer-ctc explicitly. Do not pass --backend parakeet: that is the RNN-T/TDT transducer runtime and it will reject a CTC model with "required tensor 'decoder.embed.weight' not found".

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

./build/bin/crispasr -m parakeet-ctc-1.1b-ja \
    --auto-download -f your-japanese-audio.wav

Long-form audio

For clips longer than ~15 s, prefer VAD-bounded chunking — Japanese FastConformer models drift on long single-pass windows (the safe single-pass window is ~12 s):

./build/bin/crispasr -m parakeet-ctc-1.1b-ja-q8_0.gguf \
    -f long-japanese-audio.wav --vad -t 8

Model architecture

How this was made

1. 1.The source .nemo checkpoint is the GAL checkpoint (parakeet-ja-gal.nemo) from `grider-transwithai/parakeet-ctc-1.1b-ja`. The non-GAL checkpoint in that repo has corrupt F32 weights in encoder layers 26–28 (NaN / values > 1e38) and is not usable — the GAL checkpoint is the converted one.
2. 2.Architecture hyperparameters are read from the checkpoint's model_config.yaml and cross-checked against the actual tensor shapes; the mel filterbank and Hann window are baked into the GGUF so the runtime reproduces NeMo's front-end exactly.
3. 3.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 Q80 and Q4K.
4. 4.The GGUF carries the canary-ctc architecture tag; inference runs through the shared FastConformer-CTC runtime (--backend fastconformer-ctc, auto-detected from the filename), not the RNN-T parakeet transducer backend.

Licence

Apache-2.0, inherited from the `grider-transwithai/parakeet-ctc-1.1b-ja` fine-tune. The underlying NVIDIA NeMo FastConformer-CTC architecture (`nvidia/parakeet-ctc-1.1b`) is CC-BY-4.0.

Provenance and EU AI Act Art. 53 note

• —Upstream model: grider-transwithai/parakeet-ctc-1.1b-ja — published by grider-transwithai.
• —Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• —What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• —Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository. No training-content summary was found on the upstream model card at the time of writing; that documentation gap is upstream's and is not filled here.
• —Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.