Skip to content

Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.

fix(diarization): skip orphan speaker embeddings in softened mic mode

Gabriel Stein2ffbe357d900
diff --git a/src/helpers/ipcHandlers.js b/src/helpers/ipcHandlers.js
index a3d16a1fe..d1020c0a0 100644
--- a/src/helpers/ipcHandlers.js
+++ b/src/helpers/ipcHandlers.js
@@ -7335,12 +7335,13 @@ class IPCHandlers {
// capture above.
if (!meetingSystemAudioHeard) {
if (!meetingMicDiarizationStream) {
+ const receivedAt = Date.now();
meetingMicDiarizationPath = path.join(
os.tmpdir(),
- `ow-diarize-raw-mic-${Date.now()}.pcm`
+ `ow-diarize-raw-mic-${receivedAt}.pcm`
);
meetingMicDiarizationStream = fs.createWriteStream(meetingMicDiarizationPath);
- meetingMicDiarizationStartedAt = Date.now();
+ meetingMicDiarizationStartedAt = receivedAt;
}
meetingMicDiarizationStream.write(outboundBuffer);
}
@@ -10466,10 +10467,15 @@ class IPCHandlers {
sIdx++;
}
+ // Mirrors the mic-mode single-cluster softening in mergeWithTranscript:
+ // every segment stays "you", so persisting an embedding keyed to a
+ // cluster id that owns no segments would leave the note inconsistent.
+ const micSingleClusterSoftened = diarizedSource === "mic" && speakerSet.size === 1;
+
let speakerEmbeddingsMap = null;
const speakerEmb = require("./speakerEmbeddings");
try {
- if (speakerEmb.isAvailable() && tmpWav) {
+ if (!micSingleClusterSoftened && speakerEmb.isAvailable() && tmpWav) {
const speakerIds = [...new Set(diarizationSegments.map((s) => s.speaker))];
speakerEmbeddingsMap = {};