hugging-mac/sensevoice-small-coreml

SenseVoiceSmall — Core ML for Apple Silicon

A ready-to-use Core ML conversion of FunAudioLLM/SenseVoiceSmall for fast, private speech recognition and rich audio understanding on Apple Silicon.

SenseVoiceSmall performs transcription, spoken-language identification, speech-emotion recognition, and audio-event detection in one non-autoregressive model. It is integrated into Hugging Mac for building local macOS voice apps, services, plugins, and agents.

Capabilities

• Speech recognition for Chinese, English, Cantonese, Japanese, and Korean
• Automatic spoken-language identification
• Emotion recognition: happy, sad, angry, neutral, fearful, disgusted, and surprised
• Audio-event detection including music, applause, laughter, crying, coughing, sneezing, breathing, singing, and speech noise
• Optional inverse text normalization
• Short-form audio up to 30 seconds at 16 kHz

Core ML model

The bundle contains one multifunction ML Program with three fixed feature buckets. Hugging Mac automatically selects the smallest suitable function.

Function	features input	Other inputs	logits output
encoder_100	1 × 100 × 560 FP32	lengths, language_id, style_id: 1 INT32	1 × 104 × 25055 FP16
encoder_250	1 × 250 × 560 FP32	lengths, language_id, style_id: 1 INT32	1 × 254 × 25055 FP16
encoder_500	1 × 500 × 560 FP32	lengths, language_id, style_id: 1 INT32	1 × 504 × 25055 FP16

• Architecture: non-autoregressive SANM encoder with CTC output
• Weight optimization: INT8 linear quantization
• Compute precision: FP16
• Minimum deployment target: macOS 15
• Package size: 238.7 MB
• Default function: encoder_500

The inputs are 560-dimensional acoustic features generated with the included am.mvn normalization statistics. Raw logits require CTC collapse, SentencePiece decoding, and parsing of rich language, emotion, and event tokens. The tokenizer is downloaded from the pinned upstream repository by Hugging Mac.

Use with Hugging Mac

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 SenseVoiceSmall Core ML, and use it through Live Transcription or the SpeechTranscription and SpeechUnderstanding capabilities.

For feature extraction, Core ML invocation, CTC decoding, and structured rich-transcript parsing, see the Hugging Mac SenseVoice SDK.

Provenance and integrity

• Upstream model: FunAudioLLM/SenseVoiceSmall
• Upstream revision: 3847d57b6bdf2dd8875cb1508d2af43d80a16bf7
• Conversion format: sensevoice-small-coreml
• Conversion metadata: included in sensevoice-small-coreml/conversion.json
• Directory SHA-256: 6ec1ee9247f7c717a285d809967bd64e210dbba4418e872c01a7a60bdedbb983

License

The converted model retains the FunASR Model Open Source License Agreement 1.1 and its attribution and usage conditions. Review the linked license before redistribution or commercial use. Hugging Mac is an independent open-source project and is not affiliated with or endorsed by FunAudioLLM, FunASR, or ModelScope.