saqlainnnnn/tinyvox

Purpose and entry points

A Rust workspace for a desktop voice-dictation system: capture audio, transcribe it, optionally clean/validate text, and inject resulting text into the active application. The primary executable is apps/desktop/src/main.rs; apps/desktop/src/bin/tinyvox-dict.rs is a separate dictionary-oriented CLI/binary.

Application orchestration

crates/engine/src/controller.rs is the likely central coordinator for dictation lifecycle and processing decisions. Engine state and observable events are separated into crates/engine/src/state.rs and crates/engine/src/event.rs, with public composition exposed by crates/engine/src/lib.rs.

Port-and-adapter boundary

The engine defines technology-neutral interfaces for its dependencies under crates/engine/src/ports/: audio capture (audio.rs), speech-to-text (stt.rs), cleanup (cleanup.rs), and text injection (injection.rs). This boundary lets the desktop application bind platform/service implementations without coupling controller logic to Windows APIs or particular transcription providers.

Audio and transcription pipeline

Audio acquisition is implemented as a dedicated crate at crates/audio/src/lib.rs. Speech-to-text is supplied by crates/stt-pulse/src/lib.rs, with a standalone transcription example at crates/stt-pulse/examples/transcribe.rs. The primary flow is audio samples/recording → STT transcript → engine-controlled downstream handling.

Transcript cleanup and validation

crates/cleanup/src/pipeline.rs composes transcript transformation stages, while crates/cleanup/src/validation.rs provides a validation boundary. Multiple cleanup backends are present: local/basic behavior in crates/cleanup/src/basic.rs, plus integrations named for Electron, Groq, and Llama in crates/cleanup/src/electron.rs, crates/cleanup/src/groq.rs, and crates/cleanup/src/llama.rs. These represent an important external-service/runtime boundary after transcription and before injection.

Commands, dictionaries, and text output

Voice-command interpretation is isolated in crates/voice-command/src/lib.rs, enabling recognized phrases to become actions rather than emitted prose. Dictionary domain and persistence are separated in crates/engine/src/dictionary.rs and crates/engine/src/dictionary_store.rs; the desktop dictionary binary likely manages this data. Final text delivery is performed through the engine injection port and Windows implementation at crates/win/src/injection.rs.

Platform runtime and infrastructure

The target desktop/platform layer is Windows-focused: crates/win/src/lib.rs groups active-window discovery, credential handling, text injection, and overlay support in foreground.rs, credentials.rs, and overlay.rs. Hotkey and injection examples under crates/win/examples/ indicate OS-level input integration. Rust/Cargo is the material runtime and workspace infrastructure (Cargo.toml); prototypes under prototypes/ are exploratory and outside the primary application path.