GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
cantonese-asr-lab
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
cantonese-asr-lab
Model Tree
openai/whisper-small
Input Modalities
Output Modalities
Supported Functionality
The released source-alternating Whisper-small checkpoint achieves 85.47% tolerance-two accuracy and 8.59% CER on the validation set, and 89.42% tolerance-two accuracy and 8.13% CER on the task-provided local test set.
| Evaluation surface | Items | tol2 accuracy (%) | CER (%) |
|---|---|---|---|
| Fixed validation set | 702 | 85.47 | 8.59 |
| Task-provided local test set | 1,900 | 89.42 | 8.13 |
| OOD set | 2,000 | 33.15 | 34.04 |
Metrics follow the task-defined character-level scoring protocol. The fixed validation set controls checkpoint selection. The local-test and OOD sets provide additional task-aligned and distribution-shift evaluation; they are not used to choose checkpoints, language-model weights, or fusion methods.
The task-provided Cantonese materials were distributed for the preliminary ASR task of the AI Dimsum Cup, whose task page points to the Cantonese Life Scenarios Corpus. The current dataset page is linked for provenance and is not asserted to be the exact revision used during training. WenetSpeech-Yue supplies the broader acoustic source. Training data is not included in this model repository. Information identifying the exact training-time task-data snapshot is available from the corresponding author by email; the contact address will be added with the final author record.
The checkpoint is intended for Cantonese automatic speech recognition research and reproducible evaluation with the documented normalization and decoding configuration. Users should evaluate it on their own acoustic domains and transcription conventions before deployment.
The model card does not establish suitability for high-stakes, safety-critical, or fully automated decision-making applications.
Download the complete repository because the bundled entry point loads model
assets with local_files_only=True:
bash
hf download cantonese-asr-lab/whisper-small-cantonese-w500-adaptive \--local-dir whisper-small-cantonese-w500-adaptivepython whisper-small-cantonese-w500-adaptive/predict.py \--model_dir whisper-small-cantonese-w500-adaptive \--audio_dir /path/to/audio \--test_list /path/to/test.jsonl \--output_jsonl predictions.jsonl
Input JSONL rows require an audio_path field. Output rows contain
audio_path and pred_text. Use --diagnostics-jsonl to save generated token
IDs alongside the prediction output.
The released defaults are language=zh, task=transcribe, num_beams=1, and
max_length=225. The bundled generation configuration also supplies
no_repeat_ngram_size=4 and repetition_penalty=1.05.
model.safetensors SHA-256:
a0f29a5a011213d5e4de34c40a02d42247255645f06e649e092d2dc495094370b4fda8ac549d37d8cac9797950636d50ca28f74d8e5e76e2479c15de9b956bc5config.json SHA-256:
53b4eb5c1c63510e9541417174df6ed709cd59e0b052492446ca1287088ee023generation_config.json SHA-256:
3f2ced827b5a4b0241c4f2f8883cf13da2b00ece23c341d663333c9b07b8de64A fixed 32-item offline audit found identical text and token outputs across the raw checkpoint, extracted release package, and bundled inference path.
The released model weights are available under the Apache License 2.0. Users remain responsible for complying with the separate terms that apply to upstream datasets, third-party models, and software dependencies.
predict.py SHA-256:
075d465a775f4ebff1f817d86ab16d3f8da0977257cdcab1870312aaa9c78546