moonshine-tiny-zh-onnx-fp16

moonshine-tiny-zh — fp16 ONNX

fp16 ONNX export of moonshine-ai/moonshine-tiny-zh, a 27M-parameter on-device Mandarin Chinese ASR model from the "Flavors of Moonshine" release.

This repo keeps only the fp16 variant (~55 MB: encoder 15.5 MB + decoder 39.1 MB) plus the tokenizer/config files, extracted from onnx-community/moonshine-tiny-zh-ONNX.

Files

• onnx/encoder_model_fp16.onnx — takes raw 16 kHz mono waveform (input_values, float32), outputs last_hidden_state
• onnx/decoder_model_fp16.onnx — autoregressive decoder (no KV cache), inputs input_ids + encoder_hidden_states, outputs logits
• tokenizer.json, config.json, generation_config.json, etc.

Usage

A minimal Python example (onnxruntime, greedy decoding) is maintained at github.com/xiaonengmiao/moonshine-tiny-zh-onnx-fp16:

git clone https://github.com/xiaonengmiao/moonshine-tiny-zh-onnx-fp16
cd moonshine-tiny-zh-onnx-fp16
pip install -r requirements.txt
python transcribe.py your_audio.wav   # auto-downloads this repo

Note: create the onnxruntime session with graph optimization disabled (ORT_DISABLE_ALL) — the fp16 export triggers a layer-norm fusion bug in ORT's optimizer during session initialization.

License

Inherits the upstream Moonshine AI license (other); see the original model card before commercial use.