Cross-platform speech recognition for Tauri 2.x. Desktop targets use whisper.cpp via whisper-rs; mobile delegates to the native OS engine (SFSpeechRecognizer on iOS, SpeechRecognizer on Android).

• One model, 99 languages — Whisper is multilingual; a single GGML model file handles English, Portuguese, Mandarin, and more
• No separate runtime to install — whisper-rs builds whisper.cpp statically; there is no .so/.dylib to ship
• Explicit model lifecycle — the host app controls when a model is downloaded; start_listening returns ModelNotInstalled instead of pulling hundreds of MB silently
• Hardware acceleration — opt-in metal / cuda / vulkan features map to the matching whisper.cpp backend

[dependencies]
tauri-plugin-stt = { version = "0.2", features = ["metal"] }  # macOS
# "cuda" for NVIDIA GPU, "vulkan" for cross-vendor GPU, omit for CPU

npm install tauri-plugin-stt-api

Register the plugin:

fn main() {
    tauri::Builder::default()
        .plugin(tauri_plugin_stt::init())
        .run(tauri::generate_context!())
        .unwrap();
}

{ "permissions": ["stt:default"] }

Files are fetched from HuggingFace (ggerganov/whisper.cpp) and stored under <app_data_dir>/whisper-models/. The active model is persisted to whisper-models/active.txt.

• list_models() → { models, active, total_disk_bytes }
• install_model(id) — downloads and emits stt://download-progress events
• remove_model(id) — deletes file; clears active marker if needed
• set_active_model(id) — sets which installed model start_listening loads
• unload_model() — drops the loaded Whisper context from memory; fails while listening or transcribing
• start_listening({ language?, max_duration? }) — begins a push-to-talk session
• stop_listening() — runs Whisper over captured audio and emits a final result
• is_available() — true only when a model is installed and ready
• get_supported_languages() — curated list of UI-facing locales
• check_permission() / request_permission() — microphone permission helpers

• stt://download-progress — { status, modelId, model, progress, downloaded?, total? }
• stt://result — { transcript, isFinal, confidence }
• stt://error / plugin:stt:error — { code, message, details? } (codes follow the SttErrorCode union, e.g. NO_SPEECH, AUDIO_ERROR)
• plugin:stt:stateChange — { state, isAvailable, language } (idle is emitted only after transcription finishes)

• Whisper is not a streaming recogniser. The plugin buffers audio during recording and runs a single inference pass on stop_listening. The UX is push-to-talk, not live transcription.
• Audio is captured at the device default rate, downmixed to mono, then decimated to 16 kHz with nearest-neighbour. Whisper is robust enough that a higher-quality resampler makes no measurable difference.
• Inference uses min(available_parallelism(), 4) threads — beyond that whisper.cpp shows diminishing returns, and we want headroom for the UI.

The mobile bridges expose the same JS API surface, but list_models returns an empty list and install_model / remove_model / set_active_model / unload_model are no-ops: the OS engine has no downloadable model concept. Use is_available to gate UI — on iOS/Android it reflects actual recogniser availability.

MIT