flutter_whisper_ggml 0.0.1

Offline speech recognition for Flutter powered by whisper.cpp.

Changelog

All notable changes to flutter_whisper_ggml are documented here. This project follows Keep a Changelog and uses Semantic Versioning.

0.0.1 - 2026-08-08

Initial beta release.

Added

• Bundled and pinned whisper.cpp v1.9.2 source tree; consumers do not need an external whisper.cpp checkout.
• Stable Dart FFI/C bridge for Android, iOS, Linux, macOS and Windows.
• Web Worker and Emscripten WebAssembly runtime with SIMD, pthreads, progress events and live transcription segments.
• Windows x64 CPU, CUDA and Vulkan build selection with runtime GGML device enumeration.
• Linux x64 CPU and Vulkan build targets.
• Android 24+ ARM64/x86_64 CPU/NEON builds with Android 15 16 KiB page support.
• iOS 13+ and macOS 10.15+ C++17 CocoaPods builds using Apple Accelerate.
• Typed catalog for all 34 GGML model variants supported by the bundled upstream download scripts.
• FlutterWhisper.loadModel and WhisperModelManager.prepare for automatic native downloads, application-support caching and Web URL loading.
• Streamed model downloads with progress, atomic .part writes, exact-size checks, SHA-256 verification and cache reuse.
• Manual GGML model loading from native paths, HTTP/asset URLs and browser blob URLs.
• WAV PCM/Float decoding, multichannel-to-mono mixing, 16 kHz resampling and strict input validation.
• Direct transcription of normalized mono 16 kHz Float32List samples.
• Automatic or explicit language selection, English translation, greedy/beam decoding, prompts, context control, offsets, duration, token timestamps, segment limits and probability/temperature thresholds.
• Silero VAD configuration, TinyDiarize controls and VAD result regions.
• Serialized model inference, cooperative cancellation, deterministic disposal, progress callbacks and incremental segment callbacks.
• Timestamped segments, token metadata, detected language and native timing information.
• Stable versioned JSON output and normalized SRT subtitle formatting.
• GGML magic validation and explicit rejection of incomplete upstream for-tests-* fixtures.
• Minimal all-platform smoke example plus a separate full showcase repository.
• Unified CI covering analysis, tests, pub.dev validation and release builds for Android, Linux CPU/Vulkan, Windows CPU, Web/WASM, iOS and macOS.
• MIT License, Apple privacy manifests, package documentation, model download catalog and troubleshooting guidance.

Design decisions

• FFmpeg is intentionally not a package dependency. Applications convert MP3, AAC, M4A, MP4 and other compressed media to WAV when needed.
• Model binaries are intentionally downloaded or supplied by the application so they do not inflate the plugin archive.
• Native file transcription is intentionally WAV-only for predictable audio preprocessing.

Known limitations

• The API is beta and may change before 1.0.0.
• Metal, Android Vulkan, WebGPU, microphone streaming and plugin-managed browser model persistence are not available in this release.
• Apple targets are compiled in CI but still require physical-device validation.