Moonshine Tiny · OpenASR

Tiny 27M-parameter English ASR built for real-time, on-device transcription

Native speech-to-text in the OpenASR runtime — engineered for peak performance on CPU & GPU, no Python at inference time.

---

✨ Highlights

• 🪶 Just 27M parameters — the smallest Moonshine, sized for memory- and compute-constrained edge hardware
• ⚡ Real-time on-device — engineered by Useful Sensors for live transcription and voice commands on low-cost devices
• 🎯 Accurate for its size — beats similarly-sized ASR systems on standard English benchmarks (per the Moonshine paper)
• 🗣️ English speech-to-text — sequence-to-sequence ASR trained on 200K hours of audio
• 🦀 Native in OpenASR — .oasr packs run with no Python at inference, engineered for peak performance on CPU & GPU

🚀 Quickstart

# 1. Install the OpenASR CLI  ·  https://openasr.org
# 2. Pull a build (pick a quant — see the table below)
openasr pull moonshine-tiny:q8

# 3. Transcribe
openasr transcribe audio.wav --model moonshine-tiny

All builds for this model:

openasr pull moonshine-tiny:fp16
openasr pull moonshine-tiny:q8

📦 Available builds

_{RTF = real-time factor on the fixed 11s JFK clip (lower is faster); RAM peak measured per pack in an isolated subprocess. JFK ΔWER compares each quantized build's JFK transcript to this model's fp16 JFK transcript, so it measures quantization drift rather than absolute recognition accuracy. q8_0 is the recommended default — near-reference quality at a fraction of the footprint.}

🧠 About Moonshine Tiny

Moonshine Tiny is the smallest model in Useful Sensors' Moonshine family — a 27M-parameter, sequence-to-sequence English speech-recognition model designed for real-time, on-device transcription on hardware that is severely constrained in memory and compute. Trained on 200,000 hours of audio, it transcribes English speech to text and, despite its size, reports greater accuracy than existing ASR systems of comparable scale on standard benchmarks. It targets developers building live transcription and voice-command experiences on low-cost devices. Like other autoregressive ASR models it can occasionally hallucinate or repeat on very short or clipped segments, so robust in-domain evaluation is recommended before deployment. This OpenASR repo repackages the original weights as .oasr packs that run natively in the OpenASR runtime — no Python at inference time. The q8_0 build is the recommended default (near-reference accuracy at roughly a third of the footprint); fp16 is for verification or maximum fidelity.

⚙️ How these packs were made

Converted from UsefulSensors/moonshine-tiny with the OpenASR importer:

openasr model-pack import moonshine <src> <out>.oasr \
  --package-id moonshine-tiny --quantization {fp16,q8-0,q4-k}

The .oasr container is GGUF-backed; packs use zero-copy mmap weight binding and graph buffer reuse to keep peak memory low.

⚖️ License

These packs inherit the upstream model's license: MIT (source). OpenASR packaging retains the upstream copyright and NOTICE; the only modifications are format conversion and quantization.

🙏 Acknowledgements

This pack is a redistribution of Moonshine Tiny, created and open-sourced by Useful Sensors (UsefulSensors/moonshine-tiny). All credit for the original architecture, training, and weights belongs to them; the license is inherited from and identical to the upstream model (MIT). Thank you to the Moonshine authors — Nat Jeffries, Evan King, Manjunath Kudlur, Guy Nicholson, James Wang, and Pete Warden — for releasing their work openly.

🔗 Links

• 🦀 OpenASR — https://github.com/QuintinShaw/openasr
• 🌐 Website — https://openasr.org
• 🤗 Upstream model — UsefulSensors/moonshine-tiny