fix: decode opus manually and use json audio transcription

voice.py

@@ -1,10 +1,65 @@
 from __future__ import annotations
 
+import logging
+import wave
 from pathlib import Path
 from typing import Callable
 
+from discord.opus import Decoder, OpusError
 from discord.ext import voice_recv
 
+log = logging.getLogger(__name__)
+
+
+class OpusWaveSink(voice_recv.AudioSink):
+    """Decode raw opus packets per-SSRC and write PCM to a wav file."""
+
+    def __init__(self, output_path: str):
+        super().__init__()
+        self.output_path = output_path
+        self._file: wave.Wave_write = wave.open(output_path, "wb")
+        self._file.setnchannels(Decoder.CHANNELS)
+        self._file.setsampwidth(Decoder.SAMPLE_SIZE // Decoder.CHANNELS)
+        self._file.setframerate(Decoder.SAMPLING_RATE)
+        self._decoders: dict[int, Decoder] = {}
+        self._error_counts: dict[int, int] = {}
+
+    def wants_opus(self) -> bool:
+        return True
+
+    def write(self, user, data) -> None:
+        packet = data.packet
+        ssrc = getattr(packet, "ssrc", 0)
+        decoder = self._decoders.setdefault(ssrc, Decoder())
+
+        try:
+            if packet:
+                opus_bytes = data.opus
+                if not opus_bytes:
+                    return
+                pcm = decoder.decode(opus_bytes, fec=False)
+            else:
+                pcm = decoder.decode(None, fec=False)
+        except OpusError as exc:
+            count = self._error_counts.get(ssrc, 0) + 1
+            self._error_counts[ssrc] = count
+            if count in {1, 10, 100}:
+                log.warning(
+                    "Dropping corrupted opus packet for ssrc %s: %s (count=%s)",
+                    ssrc,
+                    exc,
+                    count,
+                )
+            return
+
+        self._file.writeframes(pcm)
+
+    def cleanup(self) -> None:
+        try:
+            self._file.close()
+        except Exception:
+            log.warning("OpusWaveSink got error closing file on cleanup", exc_info=True)
+
 
 class MeetingRecorder:
     """Wrapper around discord-ext-voice-recv's listen/stop_listening API."""
@@ -13,14 +68,14 @@ class MeetingRecorder:
         self.vc = voice_client
         self.output_path = output_path
         self.recording = False
-        self.sink: voice_recv.WaveSink | None = None
+        self.sink: OpusWaveSink | None = None
 
     async def start(self, after_callback: Callable[[Exception | None], None]) -> None:
         if self.vc.is_listening():
             raise RuntimeError("Voice client is already listening")
 
         Path(self.output_path).parent.mkdir(parents=True, exist_ok=True)
-        self.sink = voice_recv.WaveSink(self.output_path)
+        self.sink = OpusWaveSink(self.output_path)
         self.vc.listen(self.sink, after=after_callback)
         self.recording = True