CoolFace
Modelpublic

betteropts/whisper-small-cantonese-ggml-q5_1

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face
0likes
Model Card

Whisper Small Cantonese GGML Q5_1 v1

This repository hosts a whisper.cpp-compatible Q5_1 GGML artifact derived from `alvanlii/whisper-small-cantonese`.

The model was converted for offline mobile use in Vocal2Script. It is intended to improve Cantonese transcription compared with a general Whisper small Q5_1 model while keeping the file size near the existing small-model download.

Files

  • —ggml-small-cantonese-q5_1-v1.bin: Q5_1 quantized whisper.cpp GGML model.
  • —ggml-small-cantonese-q5_1-v1.bin.sha256: SHA-256 checksum for the exact binary.
  • —LICENSE: Apache-2.0 license text and upstream attribution.

Artifact Metadata

  • —Source model: alvanlii/whisper-small-cantonese
  • —Source file: upstream ggml-model.bin
  • —Release filename: ggml-small-cantonese-q5_1-v1.bin
  • —Version: v1
  • —Quantization: q5_1
  • —Byte size: 190085487
  • —SHA-256: 157efd532a10c8b7d4351e46e514263bfe1c69072a525c58d3aed39b271e39c6
  • —Recommended decoder language for this v1 artifact: yue

No additional training or fine-tuning was performed for this release. The release artifact is a quantized derivative of the upstream GGML file.

Usage With whisper.cpp

sh
./build/bin/whisper-cli \
  -m ggml-small-cantonese-q5_1-v1.bin \
  -f sample.wav \
  -l yue

For environments where yue is unsupported or performs poorly, test -l zh as a fallback.

Local Smoke Test

The artifact was smoke-tested with whisper.cpp commit fc674574ca27cac59a15e5b22a09b9d9ad62aafe on macOS arm64 in CPU mode.

Validation sample:

  • —A short local Hong Kong Cantonese sample was used for validation and is not included in this repository.
  • —The sample was converted locally to 16-bit PCM, mono, 16000 Hz WAV.

Observed summary:

  • —The artifact loaded successfully with both zh and yue decoder settings.
  • —yue produced the strongest Cantonese-preserving output on the local sample.
  • —Compared with the general Whisper small Q5_1 baseline, this artifact better preserved colloquial Cantonese wording and particles.
  • —The recommended v1 decoder language is therefore yue.

Limitations

  • —This is an initial v1 artifact validated on a small local sample.
  • —Broader QA is still needed for noisy speech, long audio, silence, English-only audio, Mandarin audio, and Cantonese-English code-switching.
  • —The artifact should not be treated as production-ready until quality, runtime, memory, crash, download, and fallback behavior have been measured in the target app environment.

Attribution

This release derives from `alvanlii/whisper-small-cantonese`, which is distributed under Apache-2.0.

Please preserve upstream attribution when redistributing or integrating this artifact.