Changelog

All notable user-facing changes to Moonshine Voice are documented in this file.

The format is based on Keep a Changelog. Please keep the bullets high level, and no more than about 200 characters.

[0.1.5]

Added

• Streaming frontend graphs can load as frontend.model.ort plus frontend.weights.ort, keeping int8 weights on disk. English streaming models on quantized_26_08_21 cut the frontend download by about 75%.
• moonshine-voice[finetune] is an alias of [lora]; moonshine-voice finetune runs the same trainer.
• Built-in --dataset uwb_atcc (real VHF, research/NC), --sites encoder|both, --adapt full, and --eval-dataset atco2.
• Streaming text to speech: push text as it arrives with stream() / push_text / flush, and take audio back chunk by chunk or through an on_chunk handler.
• Streaming starts speaking part way through a sentence rather than waiting for the whole one, so the first audio arrives sooner on long replies. Piper's sub-sentence chunks are sample-exact.
• Cancelling a streamed reply now reports a cancelled status to whoever is pulling chunks, so an interrupted reply is distinguishable from one that ran out of text.
• AgentFlow.say_stream() and Dialog.say_stream() speak a language model's reply as it is generated, instead of waiting for the whole thing.
• moonshine_tts_split_utterances() exposes the sentence splitter that streaming and say() share.
• EmbeddingModel is a public low-level type in Python, JavaScript, Swift, and Java, for embedding text and scoring similarity without adopting AgentFlow.
• Japanese streaming speech-to-text in small and tiny sizes, and small streaming is now what "ja" selects by default. The older non-streaming Japanese models stay available by architecture.
• Streaming speech-to-text for German and Spanish in small and tiny, and for Arabic, Mandarin, Tagalog and Vietnamese in tiny. Streaming is now the default for each of these languages; older models stay available by architecture.
• Moonshine models are now MIT by default, in every language and at every size. Only the legacy non-streaming models for languages other than English stay under the Moonshine Community License, and that list is now enumerated in LICENSE.

Changed

• The available-models table lists streaming STT for every published language. Older Community non-streaming models that have a streaming replacement are marked deprecated.
• Streaming speech-to-text models open faster by reusing mapped .ort bytes at session create instead of copying them, matching the non-streaming path.
• Importing moonshine-voice no longer loads requests, so a CLI call against a cached model starts sooner.
• Text to speech now splits sentences with one shared implementation that keeps Dr. Smith and J. R. R. Tolkien whole and understands 。！？؟। terminators, replacing four different naive rules.
• The fp32, fp16, and q4f16 embedding model variants are no longer supported. Passing them fails with a "no longer supported" error; use q4 (the default) or q8.
• Text to speech holds the audio device open across queued utterances, so consecutive say() calls no longer leave a gap or a click between them.
• Kokoro text to speech is 4x faster on short sentences and 2x on long ones on a Raspberry Pi 4, a closer match to the reference voice, and a 10 MB smaller download. It uses about 85 MB more memory.
• Kokoro now ships as two stages, kokoro/prosody.* and kokoro/decoder.*, and no whole-utterance model. They render the same audio, so the download halves. In-memory callers should pass all four keys.
• Every Piper voice now ships as <stem>.upstream.* plus <stem>.generator.*, the same total size and the same audio, which is what lets a reply start playing before it is synthesized.
• Streamed speech is levelled from a per-voice measurement instead of being left unnormalized, so it no longer arrives several decibels quieter than say() and the same across voices.
• Domain customization docs and the fine-tune Colab no longer call ATCOSIM radio. Phraseology stays the default walkthrough; VHF is a separate command. The example lives under examples/python/finetune/.

Fixed

• moonshine_get_stt_dependencies now says 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 creating and closing one in a loop no longer retains tens of megabytes per instance.
• Kokoro text to speech failed to load in WebAssembly after the faster model landed, because the runtime was built without the quantized operators it needs.
• A native error in WebAssembly now reports its message instead of a heap address.

[0.1.3]

Added

• decode_incomplete_lines (default true). Set false to encode as audio arrives but wait until the line is complete before decoding.
• Optional moonshine-voice[lora] extra trains a decoder-only LoRA adapter on your audio (ATCOSIM example included). Default inference installs are unchanged.

Changed

• The LoRA Colab notebook 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.

Fixed

• Meeting Notes playback no longer 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.

[0.1.2] - August 13th, 2026

Added

• Runtime domain customization for streaming speech-to-text: pass keyterms to bias decoding towards jargon, or context to find the terms in a passage of text. See Domain Customization.
• Documentation has been reorganized in mkdocs style, with one file per section rather than everything in one large README.md. These docs are also available at moonshine.readthedocs.io.

Changed

• moonshine_load_transcriber_from_memory_files() rejects an unrecognized filename key with MOONSHINE_ERROR_INVALID_ARGUMENT, naming it, instead of dropping it silently and reporting the file as missing.

Fixed

• Meeting Notes no longer freezes for seconds when returning to its tab during a long recording: 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, and the wheels, archives and Android debug variant are now optimized rather than debug builds (8-15% faster streaming on a Pixel 10a).

[0.1.1] - August 6th, 2026

Added

• Speculative decoding for streaming speech-to-text, on by default. Streaming re-decodes verify the previous hypothesis and continue from the first mismatch instead of restarting from BOS, cutting live latency on several platforms (for example Medium Streaming on MacBook Pro from ~103ms to ~74ms). Disable with use_speculative_decoding=false.
• AgentFlow.otherwise() — a callback for speech that matched no trigger and no waiting prompt, so dictation-style UIs can take free-form lines without treating them as failed commands.
• Cross-platform voice cloning API: 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.

Changed

• Breaking: DialogFlow is renamed to AgentFlow everywhere (APIs, packages, and example apps).
• Breaking: High-level client APIs are regularized across Python, JavaScript, Swift, and Java around construct → configure with chainable setters → load(). In particular, Python MicTranscriber and TextToSpeech no longer download and open models inside the constructor.
• Breaking: The C++ binding (moonshine-cpp.h) follows the same higher-level shape for transcription and TTS.
• Breaking: Only OnnxRuntime flatbuffer models (.ort) are accepted. Supplying .onnx (or ONNX external-data sidecars) fails with a clear migration error. Convert with python scripts/convert-models-to-ort.py. See docs/ort-only-models.md.
• Breaking: Speaker-diarization models are no longer embedded in the library. With identify_speakers, they download on first use like other models (~8 MB off every mobile binary; Android arm64 install ~24.6 MB → ~16.4 MB, iOS linked binary ~30.6 MB → ~22.4 MB). See docs/diarization-models.md.
• Downloadable assets are served from download.moonshine.ai (and the Hugging Face mirror) instead of the old GCP-hosted buckets; 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.

Removed

• Remaining public references to the old Intent API (use AgentFlow phrase matching instead).
• Embedded diarization model payloads from the core library binary.

Fixed

• Windows MSVC path handling for OnnxRuntime sessions and file opens.
• Clearer errors when deprecated or unsupported model formats are supplied.