sayd-core

The engine behind sayd: utterance queue, sentence chunking, "write for the ear" text cleanup, configuration, and the state machine that drives them.

Deliberately depends on neither an inference stack nor an audio backend nor D-Bus. Synthesis reaches it through the Synthesizer trait and audio through AudioSink, which is what lets the whole engine be driven in a unit test with no model, no audio device and no display.

• engine — commands in, immutable snapshots out; tick() does one unit of work and returns
• handle — runs the engine on its own thread
• queue — four submission policies with per-source defaults
• chunk — two-phase: split to a character target, then re-split whatever overruns the model's phoneme-token budget once phonemised
• cleanup — URLs, markdown, code fences, acronyms
• config — TOML under XDG paths, atomic writes

Licence

MIT.