alt Hacker NewsShow HN: Moonshine Open-Weights STT models – higher accuracy than WhisperLargev3
I wanted to share our new speech to text model, and the library to use them effectively. We're a small startup (six people, sub-$100k monthly GPU budget) so I'm proud of the work the team has done to create streaming STT models with lower word-error rates than OpenAI's largest Whisper model. Admittedly Large v3 is a couple of years old, but we're near the top the HF OpenASR leaderboard, even up against Nvidia's Parakeet family. Anyway, I'd love to get feedback on the models and software, and hear about what people might build with it.
Comments
According to the OpenASR Leaderboard [1], looks like Parakeet V2/V3 and Canary-Qwen (a Qwen finetune) handily beat Moonshine. All 3 models are open, but Parakeet is the smallest of the 3. I use Parakeet V3 with Handy and it works great locally for me.
[1]: https://huggingface.co/spaces/hf-audio/open_asr_leaderboard
Claiming higher accuracy than Whisper Large v3 is a bold opening move. Does your evaluation account for Whisper's notorious hallucination loops during silences (the classic 'Thank you for watching!'), or is this purely based on WER on clean datasets? Also, what's the VRAM footprint for edge deployments? If it fits on a standard 8GB Mac without quantization tricks, this is huge.
> Models for other languages are released under the Moonshine Community License, which is a non-commercial license.
Weird to only release English as open weights.
I've helped many Twitch streamers set up https://github.com/royshil/obs-localvocal to plug transcription & translation into their streams, mainly for German audio to English subtitles.
I'd love a faster and more accurate option than Whisper, but streamers need something off-the-shelf they can install in their pipeline, like an OBS plugin which can just grab the audio from their OBS audio sources.
I see a couple obvious problems: this doesn't seem to support translation which is unfortunate, that's pretty key for this usecase. Also it only supports one language at a time, which is problematic with how streamers will frequently code-switch while talking to their chat in different languages or on Discord with their gameplay partners. Maybe such a plugin would be able to detect which language is spoken and route to one or the other model as needed?
Nice work. One metric I’d really like to see for streaming use cases is partial stability, not just final WER.
For voice agents, the painful failure mode is partials getting rewritten every few hundred ms. If you can share it, metrics like median first-token latency, real-time factor, and "% partial tokens revised after 1s / 3s" on noisy far-field audio would make comparisons much more actionable.
If those numbers look good, this seems very promising for local assistant pipelines.
Any plans regarding JavaScript support in the browser?
There was an issue with a demo but it's missing now. I can't recall for sure but I think I got it working locally myself too but then found it broke unexpectedly and I didn't manage to find out why.
Accuracy is often presumed to be english, which is fine, but it's a vague thing to say "higher" because does it mean higher in English only? Higher in some subset of languages? Which ones?
The minimum useful data for this stuff is a small table of language | WER for dataset
This is awesome, well done guys, I’m gonna try it as my ASR component on the local voice assistant I’ve been building https://github.com/acatovic/ova. The tiny streaming latencies you show look insane
No idea why 'sudo pip install --break-system-packages moonshine-voice' is the recommended way to install on raspi?
The authors do acknowledge this though and give a slightly too complex way to do this with uv in an example project (FYI, you dont need to source anything if you use uv run)
Very cool. Anyway to run this in Web assembly, I have a project in mind
For those wondering about the language support, currently English, Arabic, Japanese, Korean, Mandarin, Spanish, Ukrainian, Vietnamese are available (most in Base size = 58M params)
Implemented this to transcribe voice chat in a project and the streaming accuracy in English on this was unusable, even with the medium streaming model.
haven't tested yet but I'm wondering how it will behave when talking about many IT jargon and tech acronyms. For those reason I had to mostly run LLM after STT but that was slowing done parakeet inference. Otherwise had problems to detect properly sometimes when talking about e.g. about CoreML, int8, fp16, half float, ARKit, AVFoundation, ONNX etc.
Do you also support timestamps the detected word or even down to characters?
Streaming transcription is crazy fast on an M1. Would be great to use this as a local option versus Wispr Flow.
How does this compare to Parakeet, which runs wonderfully on CPU?
onnx models for browser possible?
fyi the typepad link in your bio is broken
If only it did Doric
How does it compare to Microsoft VibeVoice ASR https://news.ycombinator.com/item?id=46732776 ?
The streaming latency numbers are what stand out to me here. Accuracy benchmarks get all the attention, but for real-time applications (voice assistants, live captioning, in-call transcription), the tail latency matters more than shaving a few points off WER. A 58M param model that can stream with sub-second latency on a Raspberry Pi opens up a whole class of edge applications that just aren't practical with larger models, even if those larger models score higher on static benchmarks.
Very exciting stuff!
My startup is making software for firefighters to use during missions on tablets, excited to see (when I get the time) if we can use this as a keyboard alternative on the device. It's a use case where avoiding "clunky" is important and a perfect usecase for speech-to-text.Due to the sector being increasingly worried about "hybrid threats" we try to rely on the cloud as little as possible and run things either on device or with the possibility of being self-hosted/on-premise. I really like the direction your company is going in in this respect.
We'd probably need custom training -- we need Norwegian, and there's some lingo, e.g., "bravo one two" should become "B-1.2". While that can perhaps also be done with simple post-processing rules, we would also probably want such examples in training for improved recognition? Have no VC funding, but looking forward to getting some income so that we can send some of it in your direction :)