Skip to content
Search

Moonshine

lastmod 2026-03-01created 2026-03-012 edits

Open-source fast, high-accuracy speech recognition (ASR) for edge devices

About

  • A real-time speech recognition framework that does all processing on-device
  • Its streaming-based model architecture generates text in real time even while the user is speaking, and achieves a lower error rate (WER 6.65%) than Whisper Large v3
  • Runs with the same API across many platforms, including Python, iOS, Android, MacOS, Linux, Windows, and Raspberry Pi, optimized on a C++ core and OnnxRuntime
  • Includes language-specific models (English, Korean, Japanese, Spanish, etc.) and command recognition (Intent Recognition), so developers can easily build voice interfaces
  • Improves on Whisper's fixed 30-second input, lack of caching, and limited language accuracy, drawing attention as an alternative well suited to low-latency voice interfaces in edge environments

Moonshine Voice overview

  • Moonshine Voice is an open-source AI toolkit for building real-time voice applications
    • All computation runs on the local device, ensuring fast responses and privacy
    • Streaming processing lets text update even while the user is speaking
  • The models are trained from scratch based on in-house research and offer higher accuracy than Whisper Large v3
  • Available in a range of sizes, from an ultra-small 26MB model to a mid-sized 245M-parameter model
  • Multilingual support, including English, Korean, Japanese, Chinese, Spanish, Vietnamese, Arabic, and Ukrainian

Key improvements over Whisper

  • Removes Whisper's fixed 30-second input window, supporting variable-length input
  • Adds caching to cut redundant computation during streaming, greatly reducing latency
  • Training a single model per language yields higher accuracy at the same size
  • A cross-platform C++ core library allows the same API from Python, Swift, Java, and more
  • Achieves a lower error rate with a 245M-parameter model, smaller than Whisper Large v3 (1.5B parameters)

Key features and API structure

  • Integrates the speech recognition pipeline into a single library, handling microphone input, voice activity detection (VAD), transcription, speaker identification, and command recognition together
  • Core classes:
    • Transcriber: converts audio input to text
    • MicTranscriber: handles microphone input automatically
    • IntentRecognizer: natural-language command recognition
  • Event-driven design that detects state changes such as LineStarted / LineUpdated / LineCompleted in real time

Models and performance

  • Moonshine Medium Streaming (245M): WER 6.65%, better than Whisper Large v3 (7.44%)
  • Moonshine Small Streaming (123M): WER 7.84%
  • Moonshine Tiny Streaming (34M): WER 12.00%
  • The Korean Tiny model was evaluated at WER 6.46%
  • All models are provided in the OnnxRuntime-based .ort format and slimmed down with 8-bit quantization

Development and distribution

  • Installable in major environments such as Python (pip install moonshine-voice), Swift (SPM), Android (Maven), and Windows (C++ headers)
  • Provides a package optimized for Raspberry Pi, enabling real-time recognition with a USB microphone
  • Released under the MIT license (English models) and the Moonshine Community License (other language models)
  • Future roadmap: smaller mobile binaries, more languages, improved speaker identification, domain customization

Benchmarks and use cases

  • Processes more than 5x faster than Whisper, making it suitable for real-time voice interfaces
  • Designed for a target response latency of 200ms or less, usable in conversational applications
  • The command recognition example can recognize natural-language variations of commands such as “Turn on the lights”
  • Public performance validated on the HuggingFace OpenASR Leaderboard

See also

Favorite site

author Zer0 sitemap