cohere-transcribe-ja-v0.1-GGUF

cohere-asr-ja-v0.1 - GGUF

GGUF conversion of efwkjn/cohere-asr-ja-v0.1, a Japanese fine-tune of CohereLabs/cohere-transcribe-03-2026.

These files target the Cohere Transcribe runtime in CrispASR. The GGUF architecture metadata is cohere-transcribe.

Files

Usage

./build/bin/crispasr \
  --backend cohere \
  -m cohere-asr-ja-v0.1-q4_k.gguf \
  -f audio-ja.wav \
  -l ja

For lower memory startup, CrispASR can mmap GGUF weights:

CRISPASR_GGUF_MMAP=1 ./build/bin/crispasr --backend cohere \
  -m cohere-asr-ja-v0.1-q4_k.gguf -f audio-ja.wav -l ja

Conversion

Converted with CrispASR's Cohere exporter:

python models/convert-cohere-asr-to-gguf.py \
  --model-dir /path/to/efwkjn/cohere-asr-ja-v0.1 \
  --output cohere-asr-ja-v0.1.gguf

./build/bin/crispasr-quantize cohere-asr-ja-v0.1.gguf cohere-asr-ja-v0.1-q4_k.gguf q4_k

The variant set mirrors cstr/cohere-transcribe-03-2026-GGUF: F16 plus Q8_0, Q6_K, Q5_1, Q5_0, and Q4_K.

Smoke Test

All six files were smoke-tested locally with CrispASR on a Japanese WAV. Output for the sample was:

おはよう。今日も一日頑張ってね

License

The base Cohere Transcribe model is Apache-2.0. The fine-tune repository does not declare an explicit license field at the time of conversion; check the source model card before redistribution or commercial use.