Moonshine Base — the open-source English ASR model from Useful Sensors — is now available in EveryScribe's Private Transcriber. If your audio is in English and you want the fastest possible local transcription, Moonshine is the model to reach for.
What Is Moonshine?
Moonshine is a family of speech recognition models built from the ground up for on-device, real-time English transcription. Developed by Useful Sensors, Moonshine was designed to run on everything from Raspberry Pi to standard laptops — and now, via EveryScribe, directly inside your web browser.
The Moonshine Base variant we ship is the sweet spot for browser use: 61 million parameters, approximately 275 MB after quantization, and fast enough to transcribe audio faster than it plays back on most modern hardware.
What Makes Moonshine Different from Whisper?
OpenAI's Whisper is the dominant English ASR model — but it was designed for batch processing of pre-recorded audio. It operates on a fixed 30-second window, which means even a 3-second utterance gets padded to 30 seconds before processing. That wasted compute adds up.
Moonshine solves this with a variable-length audio window: it processes exactly as much audio as is given, with no padding overhead. The result is a model that's up to 5× faster than comparable Whisper variants in real-world conditions.
| Model | Parameters | Speed vs. Whisper Base |
|---|---|---|
| Whisper Tiny | 39M | baseline |
| Whisper Base | 74M | — |
| Moonshine Tiny | 27M | ~5× faster |
| Moonshine Base | 61M | ~5× faster |
Beyond speed, Moonshine uses Rotary Position Embedding (RoPE) instead of fixed sinusoidal embeddings, which gives it better generalization on variable-length inputs and streaming audio.
Accuracy: Matches Whisper Base, Beats It in Streaming Scenarios
For pre-recorded English audio, Moonshine Base matches or exceeds Whisper Base on standard benchmarks. For streaming and real-time transcription — the use cases Moonshine was actually designed for — the gap widens further in Moonshine's favor due to the elimination of the 30-second padding penalty.
The model was specifically evaluated on English speech from diverse conditions, including different accents, recording environments, and noise levels. It's robust enough for:
- Conference call recordings
- Interview and podcast transcription
- Voice memo and lecture notes
- Live streaming captions
275 MB, Runs Entirely in Your Browser
On EveryScribe, Moonshine Base is delivered as a quantized INT8 ONNX model running via WebAssembly. After a one-time download and browser cache, it operates fully offline. Your audio never touches a server.
For English-speaking users who need fast, private transcription without the overhead of larger multilingual models, Moonshine Base is the most efficient option we offer.
When to Choose Moonshine
Pick Moonshine Base when:
- Your audio is English only
- You want the smallest memory footprint among our English-capable models (275 MB)
- Speed is the priority
- You're transcribing content where every second matters — live captions, voice memos, rapid meeting notes
If you need a larger language range, check out Parakeet TDT 0.6B for English + European languages, or SenseVoice for East Asian language support.
Get Started
Visit everyscribe.com/dashboard/offline-transcriber, select Moonshine Base from the ASR model dropdown, and download it once. All future transcriptions run locally, with no network required.
Moonshine is open-source under the MIT license, maintained by Useful Sensors on GitHub and available on Hugging Face. We'd like to thank the Useful Sensors team for building a model that takes edge inference seriously.
