moonshine-streaming-small-ONNX

ONNX export of UsefulSensors/moonshine-streaming-small (Moonshine v2, the moonshine_streaming architecture) for in-browser speech-to-text with transformers.js. No official ONNX export of this architecture exists upstream; this one was built and validated by Workmind. The export is batch/full-utterance (encoder + merged decoder with past-KV), not chunked-streaming.

Usage (transformers.js)

import { pipeline } from '@huggingface/transformers';

const transcriber = await pipeline(
  'automatic-speech-recognition',
  'Workmind/moonshine-streaming-small-ONNX',
  { dtype: { encoder_model: 'q8', decoder_model_merged: 'q8' } },
);

const { text } = await transcriber(audio); // Float32Array, 16 kHz mono

Input length caveat: the v2 audio frontend reshapes input into 5 ms frames (80 samples @ 16 kHz) and fails on partial frames. transformers.js's generic feature extractor does not pad for you, so zero-pad each utterance to a multiple of 80 samples before calling the pipeline.

Compatibility note (why config.json says moonshine)

transformers.js has no moonshine_streaming registration yet, so the published config.json intentionally declares model_type: "moonshine" with v1-shaped normalized-config keys. The stock transformers.js bundle then loads this model through its registered Moonshine v1 path; all v2 architectural differences (including the sliding-window encoder attention mask, synthesized all-ones inside the graph) are baked into the ONNX graphs, which are v1-compatible in I/O. Once native moonshine_streaming support ships upstream, a new revision of this repo will carry the real config.

Files

Variant	Encoder	Decoder (merged)	Notes
fp32	encoder_model.onnx (196 MB)	decoder_model_merged.onnx (340 MB)	reference
q8	encoder_model_quantized.onnx (71 MB)	decoder_model_merged_quantized.onnx (87 MB)	recommended; transcript-lossless vs fp32
q4	encoder_model_q4.onnx (56 MB)	decoder_model_merged_q4.onnx (116 MB)	transcript-lossless vs fp32

The q8 quantization keeps the causal-Conv audio frontend weights in fp32 (int8 Conv quantization silently produces empty transcripts).

Validation

fp32 greedy decode via onnxruntime matches the PyTorch reference token-for-token at the pinned revisions. In-browser (transformers.js 4.2.0 / onnxruntime-web WASM, M-series MacBook, 11 s clip): q8/q8 loads in ~2.3 s and transcribes in ~0.9 s with an exact transcript match against the Python reference.

Provenance

Built from UsefulSensors/moonshine-streaming-small at revision 2c036506f23a09c18df5a50057599ba6d9280999 with transformers 5.14.1 and a patched optimum-onnx export config. The full reproducible pipeline (export, quantization, parity checks, config rewrite) lives at WorkmindAI/stt-models.

License and attribution

MIT, same as the upstream Moonshine weights by UsefulSensors / Moonshine AI. This repo is a converted/quantized redistribution of those weights with a compatibility-rewritten config; all credit for the model itself goes to the original authors.