| --- |
| license: cc-by-nc-4.0 |
| library_name: coremltools |
| pipeline_tag: automatic-speech-recognition |
| base_model: Audio8/Audio8-ASR-0.1B |
| language: |
| - en |
| - zh |
| - fr |
| - de |
| - ja |
| - ko |
| - yue |
| tags: |
| - coreml |
| - apple-silicon |
| - automatic-speech-recognition |
| - multilingual |
| - macos |
| --- |
| |
| # Audio8-ASR 0.1B — Core ML for Apple Silicon |
|
|
| A ready-to-use hybrid Core ML conversion of [Audio8-ASR-0.1B](https://huggingface.co/Audio8/Audio8-ASR-0.1B) for fast, private, multilingual speech transcription on Apple Silicon. |
|
|
| This model is integrated into [Hugging Mac](https://github.com/devilyouwei/hugging-mac), where you can run it locally and build macOS voice apps, transcription services, plugins, and agents. |
|
|
| > **Non-commercial license:** the model is provided under CC-BY-NC-4.0 and may not be used commercially without separate permission from the upstream rights holder. |
|
|
| ## Architecture |
|
|
| This is an explicitly hybrid runtime rather than a fully Core ML autoregressive model: |
|
|
| - The audio encoder and MLP tower run with Core ML on the Apple Neural Engine. |
| - The cached Qwen2 decoder runs with PyTorch MPS on the GPU. |
| - A multifunction ML Program provides optimized 5, 10, and 30-second audio buckets. |
| - Audio-tower weights use INT8 linear quantization with FP16 compute. |
|
|
| The package includes the Core ML audio tower, decoder and projector weights, processor configuration, and reproducible conversion metadata. The tokenizer is downloaded from the pinned upstream model by Hugging Mac. |
|
|
| ## Model details |
|
|
| | Property | Value | |
| |---|---| |
| | Languages | Chinese, English, French, German, Japanese, Korean, Cantonese | |
| | Audio | 16 kHz mono, up to 30 seconds | |
| | Parameters | 0.324B end-to-end; 0.104B language model | |
| | Core ML target | macOS 15 or later | |
| | Package size | 429 MB | |
| | Decoding | Greedy autoregressive decoding with KV cache | |
|
|
| ## Core ML functions |
|
|
| | Function | `audios` input | `attn_mask` input | `hidden` output | |
| |---|---|---|---| |
| | `tower_5s` | `1 × 128 × 500` FP32 | `1 × 1 × 65 × 65` FP32 | `65 × 1024` FP16 | |
| | `tower_10s` | `1 × 128 × 1000` FP32 | `1 × 1 × 130 × 130` FP32 | `130 × 1024` FP16 | |
| | `tower_30s` | `1 × 128 × 3000` FP32 | `1 × 1 × 390 × 390` FP32 | `390 × 1024` FP16 | |
|
|
| Inputs are 128-bin log-Mel features plus an additive attention mask. The default Core ML function is `tower_30s`; Hugging Mac automatically selects the smallest suitable bucket. |
|
|
| ## Use with Hugging Mac |
|
|
| Install the ASR dependencies and run the Hugging Mac application: |
|
|
| ```bash |
| git clone https://github.com/devilyouwei/hugging-mac.git |
| cd hugging-mac |
| uv sync --all-packages |
| uv run hugging-mac-web |
| ``` |
|
|
| Open the Models page, install **Audio8-ASR 0.1B Core ML**, then use Live Transcription or build against the `SpeechTranscription` capability. |
|
|
| For the full runtime API and implementation, see the [Audio8-ASR SDK package](https://github.com/devilyouwei/hugging-mac/tree/main/packages/hugging_mac_sdk/src/hugging_mac_sdk/models/audio8_asr). |
|
|
| ## Provenance |
|
|
| - Upstream model: [Audio8/Audio8-ASR-0.1B](https://huggingface.co/Audio8/Audio8-ASR-0.1B) |
| - Upstream revision: `8487da63d581fa4fc9b5c60444cb57c3a523d7aa` |
| - Conversion format: `audio8-asr-coreml-hybrid` |
| - Conversion metadata: included in `audio8-asr-coreml/conversion.json` |
| - No upstream remote Python code is included or executed by Hugging Mac. |
|
|
| ## Integrity |
|
|
| Directory SHA-256: |
|
|
| ```text |
| 6721e90590aee02b6091ec79e3ea451ae1997399e18fcf451f079535a44fce48 |
| ``` |
|
|
| ## License and attribution |
|
|
| The converted model retains the upstream **CC-BY-NC-4.0** license, attribution requirements, and non-commercial restriction. Hugging Mac is an independent open-source project and is not affiliated with or endorsed by Audio8 or AutoArk. |
|
|
