Expand description
Shared text-to-speech runtime.
Feature boundaries (see Cargo.toml):
system: no-model OS voice backend (sayon macOS, SAPI on Windows) with a word-timed manifest that can be passed torightkit-transcript.client: QTS worker client (spawn, chunking, model manifest verification).engine: the vendored Qwen3-TTS worker (worker, binaryrightkit-tts-worker); impliesclient.
wav and the sample-rate constants are always available.
Re-exports§
pub use client::QtsClient;pub use client::QtsError;pub use client::SynthOptions;pub use client::SynthResult;pub use mode_location::ModeLocation;pub use protocol::Mode;pub use protocol::PRESET_SPEAKERS;pub use system::chunk_sentences;pub use system::SentenceChunk;pub use system::SystemTts;pub use system::SystemTtsError;pub use system::SystemTtsManifest;pub use system::SystemTtsOptions;pub use system::SystemTtsResult;pub use system::SystemTtsSegment;
Modules§
- chunk
- Long-text chunking. QTS generates one utterance per call; long scripts are split at sentence boundaries (then clause, then word) and the audio is joined with short pauses.
- client
- Supervises the warm
rightkit-tts-workerprocess and drives synthesis. - mode_
location - Where each model mode lives, and the per-mode engine set a worker keeps.
- models
- Model files, manifest verification, and where things live.
- protocol
- Worker wire protocol: one JSON object per line. Worker replies are
prefixed with
REPLY_PREFIXso ggml/ORT noise on stdout cannot corrupt the stream; the client ignores every other line. - system
- OS speech backend. It renders to a temporary file, never opens speakers.
- wav
- Minimal mono PCM16 WAV helpers.
- worker_
args - Command line of
rightkit-tts-worker. Parsing lives in the library so it is tested under the default features; the binary (featureengine) only applies defaults and runs the worker.