Encoder: prefer static INT8 (QLinearConv); INT4 fallback. Recovers 92.7%-class quality in browser

mega-asr.js

@@ -232,21 +232,23 @@ async function loadAll() {
   setProgress(30);
 
   // 3. ONNX sessions
-  // Try INT8 encoder first (highest quality in browser); fall back to INT4
-  // encoder if the browser ORT lacks the ConvInteger op (which it does in
-  // both WebGPU and WASM backends as of 1.20).
-  setLoaderStatus("audio encoder (INT8 → INT4 fallback) ...");
-  let encLabel = "audio_encoder INT8";
-  let encUrl = `${HF_ROOT}/onnx/audio_encoder_int8.onnx`;
+  // Audio encoder: try static INT8 (QLinearConv/QLinearMatMul, browser-OK),
+  // then fall back to INT4 (MatMulNBits) if static INT8 also has issues.
+  // The OLD audio_encoder_int8.onnx uses ConvInteger (dynamic quant) which
+  // onnxruntime-web does NOT support — skipping that.
+  setLoaderStatus("audio encoder (INT8 static → INT4 fallback) ...");
   try {
-    state.encoder = await createSessionSimple(encUrl, encLabel, p => setProgress(30 + p * 10));
+    state.encoder = await createSessionSimple(
+      `${HF_ROOT}/onnx/audio_encoder_int8_static.onnx`,
+      "audio_encoder INT8 (static)",
+      p => setProgress(30 + p * 10),
+    );
   } catch (e) {
-    log(`INT8 encoder unsupported in browser (${e.message}); falling back to INT4 + sidecar`);
-    encLabel = "audio_encoder INT4";
+    log(`static INT8 encoder unsupported (${e.message}); falling back to INT4`);
     state.encoder = await createSession(
       `${HF_ROOT}/onnx/audio_encoder_int4.onnx`,
       `${HF_ROOT}/onnx/audio_encoder_int4.onnx.data`,
-      encLabel,
+      "audio_encoder INT4",
       p => setProgress(30 + p * 10),
     );
   }