Implement speedups for transcription.

src/worker/inference.worker.ts

@@ -38,6 +38,31 @@ if (env.backends.onnx.wasm) {
   env.backends.onnx.wasm.wasmPaths = "./ort/";
 }
 
+/**
+ * Set the WASM thread count for the device we are about to create a pipeline
+ * on (D-023). The server sends COOP/COEP, so SharedArrayBuffer exists and
+ * threads genuinely run — which cuts both ways:
+ * - Pure WASM inference scales well past onnxruntime-web's 4-thread default,
+ *   so use most of the machine (measured 52 s → 9.5 s on a 14 s clip).
+ * - The JSEP-based accelerator EPs (WebGPU, WebNN) are NOT thread-safe:
+ *   with numThreads > 1 (even ort's own default of ~4 once SAB exists),
+ *   inference fails with "WebGPU validation failed. Buffer with '' label is
+ *   invalid". They must run the wasm side single-threaded.
+ * Caveat: onnxruntime-web initializes its runtime on the first session
+ * creation and ignores later numThreads changes, so the first attempt in
+ * this worker decides. A wasm *fallback* after a failed accelerator attempt
+ * therefore may run single-threaded — correct, just slower; the next worker
+ * respawn (cancel does one, D-018) picks the full count up again.
+ */
+function configureWasmThreads(device: InferenceDevice): void {
+  const wasm = env.backends.onnx.wasm;
+  if (!wasm) return;
+  wasm.numThreads =
+    device === "wasm" && globalThis.crossOriginIsolated
+      ? Math.max(1, Math.min((navigator.hardwareConcurrency || 4) - 1, 12))
+      : 1;
+}
+
 // `lib: ["DOM"]` types `self` as Window; give it the worker-scope shape we use.
 interface DedicatedWorkerScope {
   postMessage(message: WorkerToUiMessage): void;
@@ -153,6 +178,7 @@ async function createPipeline(
 
   let lastError: unknown;
   for (const device of attempts) {
+    configureWasmThreads(device);
     const dtype = message.dtype ?? resolveDefaultDtype(message.modelId, device);
     try {
       const asr = await pipeline("automatic-speech-recognition", message.modelId, {