Moonshine
An on-device speech toolkit under MIT built for live streaming, with speech-to-text models trained from scratch in sizes down to 1 MB, one library across Python, the browser, phones and desktops.
What it does
Moonshine Voice is "an open source AI toolkit for developers building real-time voice agents and applications". Everything runs on the device with no account or API key. The README says it is "optimized for live streaming, with low latency by doing work while the user is still talking", and that its models are "trained from scratch, from higher accuracy than Whisper Large V3 down to tiny 1MB models". That accuracy claim is the project's own.
One library covers Python, JavaScript through WebAssembly, iOS, Android, macOS, Linux, Windows and Raspberry Pi. The toolkit also covers text to speech and conversational agents; those models are licensed separately from the speech-to-text ones.
What it replaces
Moonshine replaces Deepgram and AssemblyAI for live audio on a device, in a browser or on small hardware, where a round trip to a server is too slow or not allowed.
The licence is worth reading in full. The code, apart from a third-party folder, is MIT. The LICENSE file says "Moonshine models are released under the MIT License by default, in every language and at every size", including every streaming model and every English model. The exceptions, which the file calls exhaustive, are legacy non-streaming models in Arabic, Japanese, Korean, Mandarin, Spanish, Ukrainian and Vietnamese, thirteen models in all, under the Moonshine Community License: free for research and non-commercial use, commercial use by registration, and terminating once the licensee's annual revenue passes one million US dollars. Text-to-speech and grapheme-to-phoneme models follow their own readmes.
Where it falls short: word timestamps, speaker labelling and the full language list are not in the README; they are in the documentation.
How it runs
pip install moonshine-voice, then moonshine-voice mic --language en for a live test. A C API, JavaScript/WASM and mobile SDKs are listed. GPU requirements are not stated, which matches a toolkit designed for CPUs and phones.
Who it fits
A team building a voice interface that must respond while the person is still speaking, on hardware it does not control. A team doing bulk transcription of recordings on a GPU should look at Whisper or WhisperX.
Limits the README states
None beyond the licence exceptions above.
Sources
- Moonshine README: github.com/moonshine-ai/moonshine, read 2026-09-16.
- Moonshine LICENSE, with the model licence list: LICENSE, read 2026-09-16.
More speech-to-text engines
OpenAI's general-purpose speech recognition model, released with code and weights under MIT, in six sizes from 39 million to 1.55 billion parameters, with multilingual transcription, translation to English and language detection.
A dependency-free C and C++ port of Whisper under MIT that runs on CPUs, Apple Silicon and phones, with quantised models, an HTTP server and bindings for a dozen languages.
Batched Whisper transcription under BSD-2-Clause with word-level timestamps from forced alignment and speaker labels from pyannote, the closest open-source match to a hosted transcription API's output.
NVIDIA's speech models and toolkit: Apache-2.0 code, and the Parakeet TDT 0.6B v3 model under CC-BY-4.0 with 25 European languages, automatic language detection, punctuation and word timestamps.
An offline speech recognition toolkit under Apache-2.0 with streaming, models of about 50 MB, 20+ languages, a reconfigurable vocabulary and speaker identification, from small boards to clusters.
A runtime under Apache-2.0 from the next-generation Kaldi team that runs speech to text, text to speech, speaker diarization and voice activity detection locally on CPUs, NPUs and phones, with no internet connection.
Added September 16, 2026. Found an error on this page? Write to reveneau@licheo.com and it gets fixed in the next weekly pass.
Running one of these inside your own environment, with your own data and your own security rules, is the kind of work Reveneau does. Read how a forward deployed engagement works.