perf(tts): reduce Kokoro time-to-first-audio with session warmup and threading (#342)

desktop/scripts/check-file-sizes.mjs

@@ -64,7 +64,7 @@ const overrides = new Map([
   ["src-tauri/src/huddle/stt.rs", 580], // STT pipeline + PTT edge-detection flush + PTT gating (is_speech AND ptt_active) + barge-in for VAD mode + rubato resampler + earshot VAD + sherpa-onnx transcription
   ["src-tauri/src/huddle/preprocessing.rs", 670], // TTS text preprocessing pipeline + unified split_sentences + int_to_words 0-999999 + URL trailing punctuation preservation + 23 unit tests
   ["src-tauri/src/huddle/relay_api.rs", 510], // audio relay recv task + per-peer frame counting for remote human TTS interrupt
-  ["src-tauri/src/huddle/tts.rs", 1010], // TTS pipeline + cancel/shutdown handling + apply_fades + 18 unit tests for remote interrupt mechanism
+  ["src-tauri/src/huddle/tts.rs", 1030], // TTS pipeline + session warmup + cancel/shutdown handling + apply_fades + 18 unit tests for remote interrupt mechanism
 ]);
 
 async function walkFiles(directory) {

desktop/src-tauri/src/huddle/kokoro.rs

@@ -590,6 +590,21 @@ pub fn load_text_to_speech(model_dir: &str) -> Result<KokoroTTS, String> {
         .find(|p| p.exists())
         .ok_or_else(|| format!("no model.onnx found in {model_dir}"))?;
 
+    // ── Session threading options ────────────────────────────────────────
+    //
+    // parallel_execution — runs independent graph operators concurrently.
+    //   Kokoro's graph is mostly sequential, so the benefit is modest, but
+    //   it's safe and free to enable.
+    // intra_threads(0) — lets ONNX Runtime use all available CPU cores for
+    //   parallelism within individual operators (e.g., large matmuls).
+    //
+    // NOTE: GraphOptimizationLevel::All is already the ort default — no
+    // need to set it explicitly. Accuracy-altering flags (approximate_gelu,
+    // flush_to_zero) are intentionally omitted — they need A/B audio
+    // validation before enabling for a TTS model. memory_pattern is omitted
+    // because Kokoro input lengths vary per sentence and the ORT docs warn
+    // against it for variable-size inputs.
+
     // Try CoreML first (zero binary cost — macOS system framework).
     // If the model has ops CoreML can't handle (common with quantized models),
     // the EP registers fine but commit_from_file fails. Catch that and retry
@@ -597,6 +612,10 @@ pub fn load_text_to_speech(model_dir: &str) -> Result<KokoroTTS, String> {
     let session = {
         let mut builder_with_coreml = Session::builder()
             .map_err(|e| format!("session builder: {e}"))?
+            .with_parallel_execution(true)
+            .map_err(|e| format!("parallel execution: {e}"))?
+            .with_intra_threads(0)
+            .map_err(|e| format!("intra threads: {e}"))?
             .with_execution_providers([ort::ep::CoreML::default()
                 .with_compute_units(ort::ep::coreml::ComputeUnits::All)
                 .with_model_format(ort::ep::coreml::ModelFormat::MLProgram)
@@ -617,6 +636,10 @@ pub fn load_text_to_speech(model_dir: &str) -> Result<KokoroTTS, String> {
                 // Retry without any execution providers — pure CPU.
                 Session::builder()
                     .map_err(|e| format!("session builder (CPU fallback): {e}"))?
+                    .with_parallel_execution(true)
+                    .map_err(|e| format!("parallel execution (CPU): {e}"))?
+                    .with_intra_threads(0)
+                    .map_err(|e| format!("intra threads (CPU): {e}"))?
                     .commit_from_file(&model_path)
                     .map_err(|e| format!("load model {} (CPU): {e}", model_path.display()))?
             }

desktop/src-tauri/src/huddle/tts.rs

@@ -232,6 +232,25 @@ fn tts_worker(
         }
     };
 
+    // ── 2b. Warmup inference ─────────────────────────────────────────────────
+    // The first ONNX inference on any session is significantly slower than
+    // subsequent ones — it triggers JIT compilation, memory pool allocation,
+    // and (on CoreML) lazy model compilation. Run a short dummy synthesis and
+    // discard the output so the first real utterance runs at warm-session speed.
+    {
+        let t = std::time::Instant::now();
+        match engine.synth_chunk("warmup", "en", &style, SYNTH_STEPS, SYNTH_SPEED) {
+            Ok(_) => eprintln!(
+                "sprout-desktop: TTS warmup completed in {:.0}ms",
+                t.elapsed().as_millis()
+            ),
+            Err(e) => eprintln!(
+                "sprout-desktop: TTS warmup failed after {:.0}ms: {e} — first utterance may be slow",
+                t.elapsed().as_millis()
+            ),
+        }
+    }
+
     // ── 3. Initialise rodio output device ─────────────────────────────────────
     use rodio::Player;