Tauri Plugin STT (Speech-to-Text)

Cross-platform speech recognition plugin for Tauri 2.x. Desktop targets (Windows, macOS, Linux) use whisper.cpp via whisper-rs; mobile targets delegate to the native OS engines (SFSpeechRecognizer on iOS, SpeechRecognizer on Android).

Highlights

• One model, 99 languages — Whisper is multilingual; users download a single GGML model and it works for English, Portuguese, Mandarin, …
• No native runtime to ship — whisper-rs builds whisper.cpp statically; there is no libvosk.so / .dylib to install separately.
• Explicit model lifecycle — the host app controls when (and whether) a model is downloaded. start_listening returns ModelNotInstalled instead of silently pulling hundreds of MB.
• Hardware acceleration — opt-in metal / cuda / vulkan features map straight to the matching whisper.cpp backend.

Platform Matrix

| Platform | Engine | Model | | -------- | ----------------------------------------- | ----- | | iOS | SFSpeechRecognizer (Speech.framework) | OS | | Android | SpeechRecognizer | OS | | macOS | whisper.cpp via whisper-rs (Metal opt.) | GGML | | Windows | whisper.cpp via whisper-rs (CUDA opt.) | GGML | | Linux | whisper.cpp via whisper-rs (Vulkan opt.) | GGML |

Installation

[dependencies]
tauri-plugin-stt = { version = "0.2", features = ["metal"] } # macOS
# or "cuda" / "vulkan" — omit for plain CPU inference

Register the plugin and the four model-management commands:

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

Capability:

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

Model Catalogue

| id | display | size | tier | | ---------- | ------------ | ------ | -------------- | | tiny | Tiny | 75 MB | fastest | | base | Base | 142 MB | balanced ⭐ | | small | Small | 466 MB | accurate | | medium | Medium | 1.5 GB | very accurate | | large-v3 | Large v3 | 3.0 GB | most accurate |

Files are fetched from https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-<id>.bin and stored under <app_data_dir>/whisper-models/. The active selection is persisted to whisper-models/active.txt.

Commands

• list_models() → { models, active, total_disk_bytes }
• install_model(id) — downloads the model, emits stt://download-progress
• remove_model(id) — deletes the file and clears the active marker if needed
• set_active_model(id) — picks which installed model start_listening loads
• start_listening({ language?, max_duration? }) — push-to-talk session
• stop_listening() — runs Whisper over the captured audio and emits one final result
• is_available() — reports available: true only when a model is installed
• get_supported_languages() — curated list of UI-facing locales
• check_permission() / request_permission() — microphone permission helpers

Events

• stt://download-progress — { status, modelId, model, progress, downloaded?, total? }
• stt://result — { transcript, isFinal, confidence }
• stt://error — { code, message }
• plugin:stt:result — same payload as stt://result (legacy listener channel)
• plugin:stt:stateChange — { state, isAvailable, language }

Behaviour Notes

• Whisper is not a streaming recogniser. The plugin buffers audio while recording and runs a single pass on stop_listening. UX is push-to-talk.
• 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 high-quality resampler buys nothing measurable.
• Inference uses min(available_parallelism(), 4) threads — beyond that whisper.cpp shows diminishing returns and we want headroom for the UI.

Mobile

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

License

MIT.