transcribe_cpp 0.1.0

Offline, on-device speech-to-text for Flutter. Hands-free dictation with live text on iOS and Android: no network, no API keys, audio never leaves the phone.

Changelog

All notable changes to this package are documented here. This project follows Semantic Versioning; while the version is below 1.0.0, a 0.x.0 bump may break the API and a 0.x.y bump may not.

0.1.0 — 2026-10-12

First release: offline speech-to-text for Flutter on iOS and Android.

Added

• TranscribeSession, the entry point for transcription: open it on an installed model, then start() and stop() the microphone any number of times. state, text (settled and interim), level and capture are Live values; utterances, events and dropouts are streams that last the whole session. replay() runs a recording through the same pipeline, so settings can be compared over identical audio.
• Transcription of the app's own audio, for apps with their own capture: TranscribeSession.transcribeStream() runs a stream of chunks through the same endpointing and live text as the microphone, and TranscribeSession.transcribe() decodes a single clip. AudioFormat describes the input — any rate from 8 to 384 kHz, any channel count, 16-bit or float — and the session converts it to 16 kHz mono with a windowed-sinc resampler, flat to 0.01 dB through 4 kHz.
• ModelManager, which downloads catalogue models on request, resuming an interrupted download, sharing one transfer between callers and installing a file only once its size and SHA-256 match. installed is a Live map of each model's status and progressOf a Live download progress. On iOS models are stored in Application Support, excluded from iCloud backup.
• ModelCatalog, the transcription models transcribe.cpp publishes at 1 GB or less (58 models, 223 files), generated from the catalogue of the engine release the package links. Each ModelDescriptor carries the file's size, SHA-256 and a commit-pinned download URL. Apps can add their own entries; available hides any whose family the linked engine cannot load.
• The transcribe.cpp 0.3.1 engine (ggml/GGUF), shipped as a prebuilt dynamic xcframework with Metal acceleration on iOS. Decoding runs on a helper isolate, so it never blocks the UI isolate.
• Android (API 24+, arm64-v8a and x86_64), CPU only: the same engine release, prebuilt, with one CPU build per feature tier (dotprod, fp16, i8mm) and the fastest one the phone supports chosen at startup. Models are stored in the no-backup directory.
• Microphone capture on iOS over AVAudioEngine, surviving route changes, interruptions and media-services resets, and on Android over AudioRecord, reporting interrupted while the system silences the microphone and rebuilding after an audio-server restart. Before Android 10, which has no silencing callback, a second of exact digital silence is reported as an interruption, naming a call when one is active. Audio reaches Dart through a lock-free ring read over FFI, never across a method channel. Starting and stopping the microphone happen off the main thread, where Dart runs, so stop() never holds up the app's frames.
• Hands-free endpointing (EndpointerConfig) with hysteresis, minimum speech duration, hangover and pre/post-roll, driven by a from-scratch C port of Silero VAD whose 16 kHz weights ship as an asset (~1.2 MB), so endpointing needs no second inference runtime.
• Live text that extends rather than rewrites itself: a word is shown only once consecutive decodes agree on it, and words repeated across an utterance split at the length cap are removed.
• An optional gain stage (NormalizerConfig) targeting a fixed RMS, with peak and noise ceilings.
• Live<T>, a state type that is both a ValueListenable and a Stream replaying its current value to each new subscriber.