Add speaker-attributed meeting transcription

voice.py

@@ -6,15 +6,26 @@ from typing import Callable
 
 from discord.ext import voice_recv
 
+from recording import MANIFEST_NAME, MeetingSink
+
 logger = logging.getLogger(__name__)
 
 
 class MeetingRecorder:
-    """Wrapper around discord-ext-voice-recv's listen/stop_listening API."""
+    """Wrapper around discord-ext-voice-recv's listen/stop_listening API.
+
+    Records a hybrid capture:
+
+    * a mixed archive WAV at ``output_path`` (everyone, silence-padded)
+    * one speech-only WAV per speaker under ``tracks/`` next to the archive
+    * a ``tracks.json`` manifest in that ``tracks/`` directory
+    """
 
     def __init__(self, voice_client: voice_recv.VoiceRecvClient, output_path: str):
         self.vc = voice_client
         self.output_path = output_path
+        self.tracks_dir = str(Path(output_path).parent / "tracks")
+        self.manifest_path = str(Path(self.tracks_dir) / MANIFEST_NAME)
         self.recording = False
         self.sink: voice_recv.AudioSink | None = None
 
@@ -23,11 +34,18 @@ class MeetingRecorder:
             raise RuntimeError("Voice client is already listening")
 
         Path(self.output_path).parent.mkdir(parents=True, exist_ok=True)
-        wave_sink = voice_recv.WaveSink(self.output_path)
-        self.sink = voice_recv.SilenceGeneratorSink(wave_sink)
+        meeting_sink = MeetingSink(self.output_path, self.tracks_dir)
+        # SilenceGeneratorSink keeps the mixed archive timeline faithful when
+        # packets are dropped; MeetingSink itself excludes silence frames from
+        # the per-speaker tracks.
+        self.sink = voice_recv.SilenceGeneratorSink(meeting_sink)
         self.vc.listen(self.sink, after=after_callback)
         self.recording = True
-        logger.info("Voice receive started: output_path=%s", self.output_path)
+        logger.info(
+            "Voice receive started: output_path=%s tracks_dir=%s",
+            self.output_path,
+            self.tracks_dir,
+        )
 
     async def stop(self) -> None:
         if not self.recording: