Struct LocalTtsProvider 

pub struct LocalTtsProvider { /* private fields */ }

On-device KittenTTS via ONNX Runtime + misaki-rs G2P (no espeak / GPL).

Pool size is intentionally 1 session per model (serial inference under a mutex). Throughput concurrency is governed by ResourceGovernor TTS permits rather than unbounded parallel ONNX sessions (memory cost of multi-session pools is not free for mobile/desktop defaults).

Loaded packs participate in a weighted residency registry (JOE-1646 / JOE-1784). Default providers share the process-global pool; engines own isolated pools.

Implementations

impl LocalTtsProvider

pub fn new(cache_dir: PathBuf) -> Self

pub fn with_runtime( cache_dir: PathBuf, pool: Arc<TtsSessionPool>, governor: Arc<ResourceGovernor>, ) -> Self

Construct with an explicit TTS pool and governor (JOE-1784 / engine path).

pub fn with_progress(self, v: bool) -> Self

pub fn with_local_only(self, v: bool) -> Self

pub fn with_max_chars(self, n: usize) -> Self

pub fn pool(&self) -> &Arc<TtsSessionPool>

pub fn clear_sessions(&self)

Drop idle loaded ONNX sessions in this provider’s pool.

Active synthesis leases keep their registry entry and residency weight until the pin drops — a clear during an in-flight synth cannot make the same key reloadable as a second native session (JOE-1646).

Does not delete cached files under the TTS cache directory.

Trait Implementations

impl SynthesisProvider for LocalTtsProvider

fn name(&self) -> &'static str

fn synthesize<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, text: &'life1 str, opts: &'life2 SynthesisOptions, ) -> Pin<Box<dyn Future<Output = Result<SynthesisResult>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait,

fn preload<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, model: &'life1 str, voice: &'life2 str, ) -> Pin<Box<dyn Future<Output = Result<()>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait,