Kitsune-Transcribe: Japanese speech-to-text
Four Japanese speech recognition models made by structurally pruning Cohere-transcribe and Parakeet. All were trained on only Japanese content, thusly all the Parameter capacity only goes to Japanese.
| Model | Type | Params | M4 CER | Speed (RTFx) | Repository |
|---|---|---|---|---|---|
| T-0.6B | encoder-decoder | 617M | 9.42 % | 763x | Multy123/kitsune-transcribe-t0.6b-ja |
| P-0.3B | CTC | 309M | 9.73 % | 5,848x | Multy123/kitsune-transcribe-p0.3b-ja |
| P-0.1B | CTC | 98M | 11.45 % | 9,544x | Multy123/kitsune-transcribe-p0.1b-ja |
| P-0.05B | CTC | 52M | 16.35 % | 11,800x | Multy123/kitsune-transcribe-p0.05b-ja |
Results
Character error rate (CER, %), greedy decoding on the full test sets, lower is better. M4 is the mean of JSUT, CV8, ReazonSpeech and Galgame (neutral). Speed: RTFx (seconds of audio per second of processing), one run on one RTX 5090 (bf16, PyTorch eager, greedy) over 200 test utterances in batches of up to 400 s of audio; the Kitsune models were timed on their 1,000-hour study checkpoints of the same architectures. Speeds differ between machines (up to about 2x on a second RTX 5090 host) and are not comparable with the Open ASR Leaderboard's.
| Model | Params | JSUT | CV8 | ReazonSpeech | Galgame (neutral) | Galgame (all) | Emilia | M4 | Speed (RTFx) |
|---|---|---|---|---|---|---|---|---|---|
| Kitsune T-0.6B | 617M | 8.09 | 8.62 | 7.41 | 13.56 | 19.90 | 39.76 | 9.42 | 763x |
| Kitsune P-0.3B | 309M | 7.19 | 8.62 | 10.12 | 12.99 | 19.29 | 40.32 | 9.73 | 5,848x |
| Kitsune P-0.1B | 98M | 8.99 | 10.78 | 12.62 | 13.42 | 19.80 | 41.56 | 11.45 | 9,544x |
| Kitsune P-0.05B | 52M | 13.24 | 16.59 | 20.19 | 15.38 | 21.61 | 47.73 | 16.35 | 11,800x |
| Cohere Transcribe 03-2026 (T's teacher) | 2.07B | 8.30 | 4.07 | 6.28 | 14.75 | 21.42 | 36.96 | 8.35 | 444x |
| Parakeet TDT-CTC 0.6B ja, CTC head (P's teacher) | 611M | 6.71 | 7.58 | 9.71 | 13.32 | 19.64 | 40.61 | 9.33 | 3,989x |
| Whisper large-v3 | 1.54B | 7.19 | 8.51 | 14.50 | 12.41 | 21.90 | 30.65 | 10.65 | 119x |
| Whisper large-v3-turbo | 809M | 7.30 | 9.30 | 10.70 | 13.07 | 21.52 | 39.34 | 10.09 | 240x |
| Kotoba-Whisper v2.0 | 756M | 8.39 | 9.23 | 11.66 | 13.17 | 20.02 | 39.18 | 10.61 | 267x |
| Whisper small | 242M | 14.24 | 15.06 | 27.46 | 18.50 | 28.25 | 37.76 | 18.82 | 346x |
Using the models
Input: 16 kHz mono audio, split long audio at pauses (each card gives its model's chunk length). Each model card has transformers code for the bfloat16 root and "kitsune.quant.load_quantized" code for the quantised folders.
Training in short
Knowledge distillation from the teacher's outputs (soft targets plus its greedy transcript), the students initialised from the pruned teacher. Data: 11,533 h from Galgame_Speech_ASR, ReazonSpeech, Emilia-YODAS ja and Emolia ja, kept where the teacher agreed with a second model. Each card has its model's schedule, augmentation and provenance.