Fermion ResearchFermion Research

Phonon/facts checked 13 November 2026

Phonon-2.5 vs Moonshine

Both are small open speech models built for devices. Moonshine goes smaller and reaches iOS, the browser and the Raspberry Pi; Phonon-2.5 covers more European languages and has lower English word error on most sets.

Phonon-2.5 downloads and benchmarks

The short answer

Pick Moonshine when the model has to be as small as possible or run on iOS, in a browser or on a Raspberry Pi: its streaming models go down to 34 million parameters. Pick Phonon-2.5 for European languages (25 against Moonshine's 10 speech-to-text languages, most of them outside Europe) and for English accuracy: by Moonshine's own published numbers, Phonon-2.5 English has lower word error than Moonshine Medium Streaming on six of seven shared English sets.

Moonshine has no row in the Open ASR Leaderboard snapshot we compare against, and we have not run it, so the English comparison below uses Moonshine's own figures beside ours.

Side by side

Moonshine facts from github.com/moonshine-ai/moonshine, the moonshine-ai model cards on Hugging Face and moonshine-voice.readthedocs.io, read 2026-11-13.

Phonon-2.5Moonshine (streaming models)
Model typeOpen weights on Hugging FaceOpen weights
Size156 MB download (Phonon-2.5 English: 158 MB)Tiny 34 M, Small 123 M, Medium 245 M parameters
Languages25 European languages (Phonon-2.5 English: English only)English, Spanish, German, Mandarin, Japanese, Korean, Vietnamese, Tagalog, Ukrainian, Arabic
English word error, seven Open ASR sets5.06 % (Phonon-2.5 English: 4.82 %)No row in the snapshot; Moonshine's own per-set figures below
PriceNo per-hour fee when self-hostedFree to run; no hosted price listed
Where it runsYour hardware: CPUs (Linux x86-64 and arm64, Windows x86-64), Apple silicon, NVIDIA GPUs, AndroidPython, JavaScript/WASM, iOS, Android, macOS, Linux, Windows, Raspberry Pi
StreamingYes, live WebSocket from fermion serveYes, designed for live audio
DiarizationNoNot mentioned in Moonshine's README
Word timestampsYes (--json output)Not mentioned in Moonshine's README

English: our numbers beside Moonshine's

Moonshine publishes per-set English word error for Moonshine Medium Streaming on the Open ASR Leaderboard's test sets, measured by Moonshine on its floating-point reference. Its 6.65 % headline average covers eight sets including TED-LIUM, so we show the seven sets both report rather than an average. On those seven, Phonon-2.5 English is lower on six; Moonshine Medium is lower on SPGISpeech (2.58 % against 3.31 %). The two columns come from different runs, so read the gaps as indicative.

Word error rate in %, lower is better. Phonon: our run of the leaderboard's code on the full sets. Moonshine: its own model card.
Test setPhonon-2.5Phonon-2.5 EnglishMoonshine Medium Streaming (Moonshine's figures)
LibriSpeech clean1.601.642.08
LibriSpeech other3.413.565.00
AMI9.128.7610.68
Earnings-226.726.2211.90
GigaSpeech8.047.999.46
SPGISpeech3.313.312.58
VoxPopuli3.232.288.54
Average of the seven5.064.82not published

Where Moonshine is the better choice

  • Smallest footprint: Tiny Streaming at 34 million parameters, and Moonshine's README lists models down to about 1 MB.
  • Platforms Phonon does not ship for: iOS, the browser through WASM, and Raspberry Pi.
  • Mandarin, Japanese, Korean, Vietnamese, Tagalog and Arabic.
  • Streaming designed in: work starts while the user is still talking.

Where Phonon-2.5 is the better choice

  • 25 European languages in one model, including French, Italian, Portuguese, Dutch, Polish and the Nordic and Baltic languages that Moonshine does not list.
  • English accuracy: lower word error than Moonshine Medium on six of seven sets by Moonshine's own figures, most of all on earnings calls and VoxPopuli.
  • A server with an OpenAI-shaped file route and a WebSocket stream from one command, plus Docker images for CPU and CUDA.
pip install fermion-research
fermion transcribe phonon-2.5 interview.wav --json
fermion serve phonon-2.5   # POST /v1/audio/transcriptions, GET /v1/audio/stream

Questions

Is Phonon-2.5 more accurate than Moonshine?
On English, by Moonshine's own published per-set figures for Moonshine Medium Streaming, Phonon-2.5 English has lower word error on six of seven Open ASR Leaderboard sets; Moonshine is lower on SPGISpeech. Moonshine is not in the leaderboard snapshot we use, and we have not run it.
Which is smaller, Phonon-2.5 or Moonshine?
Moonshine. Its streaming models have 34 million (Tiny), 123 million (Small) and 245 million (Medium) parameters. Phonon-2.5 is a 156 MB download.
Does Moonshine run on iPhone or in the browser?
Yes, Moonshine's README lists iOS and JavaScript/WASM as well as Android, desktop and Raspberry Pi. Phonon runs on CPUs, Apple silicon Macs, NVIDIA GPUs and Android.
Which languages does Moonshine transcribe?
Moonshine's README lists English, Spanish, German, Mandarin, Japanese, Korean, Vietnamese, Tagalog, Ukrainian and Arabic. Phonon-2.5 covers 25 European languages.
Do both support live transcription?
Yes. Moonshine's streaming models process audio while the user is speaking; Phonon streams through the WebSocket route of fermion serve.

Related

Sources and method

  • Moonshine languages and platforms: github.com/moonshine-ai/moonshine, read 2026-11-13.
  • Moonshine model sizes and English figures: huggingface.co/moonshine-ai/moonshine-streaming-medium and moonshine-voice.readthedocs.io/en/latest/models/available-models/, read 2026-11-13 (Moonshine's own measurements).
  • Phonon-2.5 and Phonon-2.5 English: Fermion Research measurement, Phonon-2.5 release post (2026-10-22), Open ASR Leaderboard scoring code, full test sets of the same seven sets.

Canonical: https://www.fermionresearch.com/models/phonon-2-5/vs-moonshine/