Release v0.1.6
Moonshine Voice
AIOpen source toolkit for real-time voice agents: low-latency on-device speech to text, intent recognition and text to speech.
Release activity
| Aug | Sep | Oct | Nov | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Sunday | No releases on Aug 2, 2026 | No releases on Aug 9, 2026 | No releases on Aug 16, 2026 | No releases on Aug 23, 2026 | No releases on Aug 30, 2026 | No releases on Sep 6, 2026 | No releases on Sep 13, 2026 | No releases on Sep 20, 2026 | No releases on Sep 27, 2026 | No releases on Oct 4, 2026 | No releases on Oct 11, 2026 | No releases on Oct 18, 2026 | No releases on Oct 25, 2026 | No releases on Nov 1, 2026 | No releases on Nov 8, 2026 | No releases on Nov 15, 2026 |
| Monday | No releases on Aug 3, 2026 | No releases on Aug 10, 2026 | No releases on Aug 17, 2026 | 1 release on Aug 24, 2026 | No releases on Aug 31, 2026 | No releases on Sep 7, 2026 | No releases on Sep 14, 2026 | No releases on Sep 21, 2026 | No releases on Sep 28, 2026 | No releases on Oct 5, 2026 | No releases on Oct 12, 2026 | No releases on Oct 19, 2026 | No releases on Oct 26, 2026 | No releases on Nov 2, 2026 | No releases on Nov 9, 2026 | No releases on Nov 16, 2026 |
| Tuesday | No releases on Aug 4, 2026 | No releases on Aug 11, 2026 | 1 release on Aug 18, 2026 | No releases on Aug 25, 2026 | No releases on Sep 1, 2026 | No releases on Sep 8, 2026 | No releases on Sep 15, 2026 | No releases on Sep 22, 2026 | No releases on Sep 29, 2026 | No releases on Oct 6, 2026 | No releases on Oct 13, 2026 | No releases on Oct 20, 2026 | No releases on Oct 27, 2026 | No releases on Nov 3, 2026 | No releases on Nov 10, 2026 | No releases on Nov 17, 2026 |
| Wednesday | No releases on Aug 5, 2026 | No releases on Aug 12, 2026 | No releases on Aug 19, 2026 | No releases on Aug 26, 2026 | No releases on Sep 2, 2026 | No releases on Sep 9, 2026 | No releases on Sep 16, 2026 | No releases on Sep 23, 2026 | No releases on Sep 30, 2026 | No releases on Oct 7, 2026 | No releases on Oct 14, 2026 | No releases on Oct 21, 2026 | No releases on Oct 28, 2026 | No releases on Nov 4, 2026 | No releases on Nov 11, 2026 | |
| Thursday | No releases on Aug 6, 2026 | 1 release on Aug 13, 2026 | No releases on Aug 20, 2026 | No releases on Aug 27, 2026 | No releases on Sep 3, 2026 | No releases on Sep 10, 2026 | No releases on Sep 17, 2026 | No releases on Sep 24, 2026 | No releases on Oct 1, 2026 | No releases on Oct 8, 2026 | No releases on Oct 15, 2026 | No releases on Oct 22, 2026 | No releases on Oct 29, 2026 | No releases on Nov 5, 2026 | No releases on Nov 12, 2026 | |
| Friday | 1 release on Aug 7, 2026 | No releases on Aug 14, 2026 | No releases on Aug 21, 2026 | No releases on Aug 28, 2026 | No releases on Sep 4, 2026 | No releases on Sep 11, 2026 | No releases on Sep 18, 2026 | No releases on Sep 25, 2026 | No releases on Oct 2, 2026 | No releases on Oct 9, 2026 | No releases on Oct 16, 2026 | No releases on Oct 23, 2026 | No releases on Oct 30, 2026 | 1 release on Nov 6, 2026 | No releases on Nov 13, 2026 | |
| Saturday | No releases on Aug 8, 2026 | No releases on Aug 15, 2026 | No releases on Aug 22, 2026 | No releases on Aug 29, 2026 | No releases on Sep 5, 2026 | No releases on Sep 12, 2026 | No releases on Sep 19, 2026 | No releases on Sep 26, 2026 | No releases on Oct 3, 2026 | No releases on Oct 10, 2026 | No releases on Oct 17, 2026 | No releases on Oct 24, 2026 | No releases on Oct 31, 2026 | No releases on Nov 7, 2026 | No releases on Nov 14, 2026 |
11 releases since Jul 10, 2026
Changelog
- Streaming frontend graphs can load as frontend.model.ort plus frontend.weights.ort, keeping int8 weights on disk, and English streaming models on quantized_26_08_21 cut the frontend download by about 75%
- moonshine-voice[finetune] is an alias of [lora], and moonshine-voice finetune runs the same trainer
- Built-in --dataset uwb_atcc, --sites encoder|both, --adapt full, and --eval-dataset atco2 options for fine-tuning
- Streaming text to speech via stream(), push_text, and flush, delivering audio chunk by chunk or through an on_chunk handler
- Streaming text to speech starts speaking part way through a sentence, and Piper sub-sentence chunks are sample-exact
- Cancelling a streamed reply reports a cancelled status to whoever is pulling chunks, distinguishing interrupted replies from ones that ran out of text
- AgentFlow.say_stream() and Dialog.say_stream() speak a language model's reply as it is generated
- moonshine_tts_split_utterances() exposes the sentence splitter shared by streaming and say()
- EmbeddingModel is a public low-level type in Python, JavaScript, Swift, and Java for embedding text and scoring similarity without AgentFlow
- Japanese streaming speech-to-text in small and tiny sizes, with small streaming now the default for "ja"
- Streaming speech-to-text for German and Spanish in small and tiny, and for Arabic, Mandarin, Tagalog and Vietnamese in tiny, with streaming now the default for each of these languages
- Moonshine models are now MIT licensed by default in every language and size, with only legacy non-streaming models for non-English languages remaining under the Moonshine Community License, enumerated in LICENSE
- The available-models table lists streaming STT for every published language, and older Community non-streaming models with streaming replacements are marked deprecated
- Streaming speech-to-text models open faster by reusing mapped .ort bytes at session create instead of copying them
- Importing moonshine-voice no longer loads requests, so CLI calls against a cached model start sooner
- Text to speech uses one shared sentence splitter that keeps abbreviations like Dr. Smith and J. R. R. Tolkien whole and recognizes the terminators 。!?؟।
- Text to speech holds the audio device open across queued utterances, removing gaps and clicks between consecutive say() calls
- Kokoro text to speech is 4x faster on short sentences and 2x faster on long ones on a Raspberry Pi 4, matches the reference voice more closely, and has a 10 MB smaller download, using about 85 MB more memory
- Kokoro ships as two stages, kokoro/prosody.* and kokoro/decoder.*, with no whole-utterance model, halving the download while rendering the same audio; in-memory callers should pass all four keys
- Every Piper voice ships as <stem>.upstream.* plus <stem>.generator.*, with the same total size and audio, allowing a reply to start playing before it is fully synthesized
- Streamed speech is levelled from a per-voice measurement, so it is no longer several decibels quieter than say() and is consistent across voices
- Domain customization docs and the fine-tune Colab no longer reference ATCOSIM radio; phraseology remains the default walkthrough and VHF is a separate command under examples/python/finetune/
- moonshine_get_stt_dependencies now states whether the language is unknown or the architecture is unpublished and lists that language's architectures (GitHub issue #214)
- Streaming no longer logs "Memory is empty" or drops hypotheses when short chunks arrive faster than encoder lookahead, including on medium-streaming (GitHub issue #218)
- AgentFlow no longer downloads the q4 embedding model and then tries to open the fp32 file, which crashed load() (GitHub issue #210)
- Destroying a file-backed Transcriber now unmaps every .ort it opened, so repeatedly creating and closing one no longer retains tens of megabytes per instance
- Kokoro text to speech failed to load in WebAssembly because the runtime lacked the quantized operators it needs
- A native error in WebAssembly now reports its message instead of a heap address
- The fp32, fp16, and q4f16 embedding model variants are no longer supported and fail with a "no longer supported" error; use q4 (the default) or q8
- Added the decode_incomplete_lines option (default true), which when set to false encodes audio as it arrives but waits until the line is complete before decoding.
- Added an optional moonshine-voice[lora] extra that trains a decoder-only LoRA adapter on your audio, with an ATCOSIM example included; default inference installs are unchanged.
- The LoRA Colab notebook now calls the same fit_adapter and ATCOSIM helpers as python -m moonshine_voice.lora instead of inlining the trainer.
- Streaming speaker diarization analyzes at most one segmentation window per audio append (Stop still drains the rest) and skips embedding inference on silent speaker classes.
- Meeting Notes waits until a line is complete before decoding and writes each finished line on its own line in the document.
- Meeting Notes playback no longer produces clicks from dropped capture frames, resampler phase jumps, or mixing a second copy of the meeting that the microphone heard.
- C API streaming comments no longer refer to an undeclared out_transcript, a missing moonshine-test-v2.cpp, or the old transcribe_stream_chunk name.
- transcribe_stream after stop_stream now transcribes leftover audio, so a final transcript no longer requires earlier partial updates.
- Runtime domain customization for streaming speech-to-text by passing keyterms to bias decoding towards jargon, or context to find the terms in a passage of text
- Documentation reorganized in mkdocs style with one file per section rather than one large README.md
- moonshine_load_transcriber_from_memory_files() rejects an unrecognized filename key with MOONSHINE_ERROR_INVALID_ARGUMENT, naming it, instead of silently dropping it and reporting the file as missing
- Wheels, archives and the Android debug variant are now optimized rather than debug builds, giving 8-15% faster streaming on a Pixel 10a
- Meeting Notes no longer freezes for seconds when returning to its tab during a long recording, as capture audio is batched and the main thread yields while catching up
- Core library builds no longer write into the source tree, where targets clobbered each other's archives
- Speculative decoding for streaming speech-to-text, on by default, which verifies the previous hypothesis and continues from the first mismatch instead of restarting from BOS, cutting live latency; disable with use_speculative_decoding=false
- AgentFlow.otherwise() callback for speech that matched no trigger and no waiting prompt
- Cross-platform voice cloning API with cloning(), clone_from(), and start_cloning() / VoiceClone for file, PCM, or live microphone capture, aligned across Python, JavaScript, Swift, and Java
- AgentFlow and voice-cloning examples on Android and iOS/Swift, plus dictation and meeting note taker web examples
- Language tabs on the web example code snippets
- Hugging Face mirror of every downloadable asset at moonshine-ai/moonshine-voice-assets
- DialogFlow renamed to AgentFlow everywhere, including APIs, packages, and example apps
- High-level client APIs regularized across Python, JavaScript, Swift, and Java around construct, configure with chainable setters, then load()
- Python MicTranscriber and TextToSpeech no longer download and open models inside the constructor
- The C++ binding (moonshine-cpp.h) follows the same higher-level shape for transcription and TTS
- Only OnnxRuntime flatbuffer models (.ort) are accepted; supplying .onnx or ONNX external-data sidecars fails with a clear migration error
- Speaker-diarization models are no longer embedded in the library and download on first use with identify_speakers, reducing Android arm64 install size from about 24.6 MB to 16.4 MB and iOS linked binary size from about 30.6 MB to 22.4 MB
- Downloadable assets are served from download.moonshine.ai and the Hugging Face mirror instead of the old GCP-hosted buckets, and large model blobs are no longer kept in git LFS in this repo
- Smaller WebAssembly downloads after ORT-only packaging and cleanup of unused model files
- Improved quantization accuracy for shipped models
- Streaming transcription backs off the inference interval when it falls behind realtime to reduce piled-up work
- Windows MSVC path handling for OnnxRuntime sessions and file opens
- Clearer errors when deprecated or unsupported model formats are supplied
- Remaining public references to the old Intent API, in favor of AgentFlow phrase matching
- Embedded diarization model payloads from the core library binary
Release v0.1.0
Release v0.0.73
Release v0.0.71
Release v0.0.69
Release v0.0.68
Release v0.0.67
More in AI
View allOpenAI's coding agent, running locally in the terminal.
Get up and running with large language models locally.
Reverse engineer anything with agents, from app behavior down to native binaries.
Your own personal AI assistant. Any OS. Any Platform. The lobster way. 🦞
The open agent skills tool — install and manage skills for coding agents with npx skills.
Skills for Real Engineers. Straight from my .agents directory.
Discussion
If you publish Moonshine Voice, you can claim this product by proving you administer its repository.