Floure — Local Speech-to-Text

Floure listens to your mic, transcribes your speech, cleans it up with an LLM, and types it into whatever you're focused on. No cloud required. No account. The default setup runs entirely on your machine.

mic (cpal) → Silero VAD → Parakeet / Whisper (sherpa-onnx)
  → LLM cleanup (local S1-Mini via llama.cpp, or OpenRouter cloud)
  → type into focused window / clipboard

Models

Downloaded on demand from the Models page (resumable, verified — see docs/adr/0003-model-downloads.md).

Model	Backend	Size	Notes
Parakeet TDT 0.6B v2 (int8)	sherpa-onnx	~460 MB	Default ASR. Fast English dictation. Recommended.
Whisper large-v3-turbo	sherpa-onnx	~540 MB	Multilingual, high accuracy
Whisper base	sherpa-onnx	~200 MB	Lightweight, any language
Silero VAD	sherpa-onnx	~630 KB	Speech detection, auto-downloaded
S1-Mini Q4_K_M	llama.cpp	~462 MB	Default local LLM for cleanup. Recommended.
Gemma 3 1B IT Q4_K_M	llama.cpp	~806 MB	Alternative local LLM

Cloud cleanup (optional): OpenRouter (OPENROUTER_API_KEY, default openai/gpt-4o-mini).

Run from source

Prerequisites: pnpm, Rust toolchain, a mic, and (Linux) PipeWire/PulseAudio.

cd application && pnpm install
pnpm tauri dev          # full desktop app (Vite + Rust backend)

Useful slices:

cd application && pnpm dev                 # frontend only
cd application && pnpm build               # frontend build
cd application/src-tauri && cargo check    # Rust check
cd application/src-tauri && cargo test     # Rust tests
cd application && npx vitest run           # frontend tests

Local GPU offload (optional)

Default builds link llama.cpp statically with CPU inference and need no Vulkan toolchain — the result is a single self-contained binary. ASR is CPU either way (int8 is already sub-second per utterance).

To offload the local LLM to a GPU, build the Vulkan backend:

cd application && pnpm tauri build -- --features vulkan

That requires a Vulkan toolchain at build time (Vulkan loader + headers, a GPU with a Vulkan ICD, glslc and SPIRV-Headers). At runtime the device is picked automatically — discrete NVIDIA → AMD → CPU (application/src-tauri/src/compute.rs) — and can be overridden with FLOURE_COMPUTE=cpu|vulkan and FLOURE_MAIN_GPU=<index>.

Configuration

What	Where
App config	~/.config/floure/config.json
Transcripts (SQLite + FTS5)	~/.local/share/floure/history.db
Models	~/.local/share/floure/models/
Data dir override	STT_DATA_DIR env var

Copy .env.example to .env for API keys and optional overrides.

Desktop UI

• Onboarding wizard — system checks, mic setup, model download, permissions
• Live transcription feed — mic level, waveform, push-to-talk (CommandOrControl+Shift+Space)
• Model management — browse, download with progress, status, delete, disk usage
• History — full-text search over past transcripts (SQLite FTS5), recopy, favorites
• Dictionary — custom vocabulary with auto-replacement
• Settings — LLM provider/mode, API keys, ASR profile, language, hotkey
• Insights — usage heatmap, streaks, stats
• Widget mode — compact always-on-top mini window; system tray with start/stop

Transcripts are saved with raw text, cleaned text, LLM mode, provider, model, and timestamp.

Docs

• CONTEXT.md — project brief, layout, commands
• docs/adr/ — architecture decisions (native backend, DB path, model downloads)
• docs/ — algorithms and voice research notes

License

GNU General Public License v2.0 — see LICENSE.

---

_{Designed and developed by Akshat Kotpalliwar}