Try fixing bad live transcription performance.

src/worker/inference.worker.ts

@@ -209,22 +209,22 @@ async function transcribe(message: TranscribeMessage): Promise<void> {
   }
   const asr = await active.promise;
 
-  const streamer = new WhisperTextStreamer(
-    // The ASR pipeline for Whisper models always carries a WhisperTokenizer.
-    asr.tokenizer as ConstructorParameters<typeof WhisperTextStreamer>[0],
-    {
-      skip_prompt: true,
-      callback_function: (text: string) =>
-        post({ type: "partial", text, requestId: message.requestId }),
-    },
-  );
-
   const options: Record<string, unknown> = {
     chunk_length_s: 30,
     stride_length_s: 5,
     return_timestamps: true,
-    streamer,
   };
+  if (message.requestId === undefined) {
+    options.streamer = new WhisperTextStreamer(
+      // The ASR pipeline for Whisper models always carries a WhisperTokenizer.
+      asr.tokenizer as ConstructorParameters<typeof WhisperTextStreamer>[0],
+      {
+        skip_prompt: true,
+        callback_function: (text: string) =>
+          post({ type: "partial", text, requestId: message.requestId }),
+      },
+    );
+  }
   // Only meaningful for multilingual models; the UI omits these for `.en`
   // variants. "auto" / undefined lets Whisper detect the language.
   if (message.language !== undefined && message.language !== "auto") {