Skip to content

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

fix: Reconciliation deleted a manual mapping whenever the live and offline speaker ids matched

Chadpiha7e13884ffc68
diff --git a/src/helpers/liveSpeakerIdentifier.js b/src/helpers/liveSpeakerIdentifier.js
index 095498645..9252b1141 100644
--- a/src/helpers/liveSpeakerIdentifier.js
+++ b/src/helpers/liveSpeakerIdentifier.js
@@ -710,12 +710,24 @@ class LiveSpeakerIdentifier {
const matchedProfile = this._findStoredProfileMatch(embedding);
if (matchedProfile) {
speakerId = this._findTransientSpeakerForProfile(matchedProfile.id);
+ let forced = false;
if (!speakerId) {
- speakerId = this._assignOrForceCluster(embedding);
+ ({ speakerId, forced } = this._assignOrForceCluster(embedding));
} else if (updateCentroid) {
this._updateCentroid(speakerId, embedding);
}
+ // A forced at-cap fold lands on someone else's cluster — retagging it
+ // with the overflow speaker's profile would erase the original identity
+ // and keep capturing this voice via the profile lookup even after the
+ // cap rises.
+ if (forced) {
+ return {
+ speakerId,
+ displayName: this.transientDisplayNames.get(speakerId) || null,
+ };
+ }
+
this.transientProfileIds.set(speakerId, matchedProfile.id);
this.transientDisplayNames.set(speakerId, matchedProfile.display_name);
return {
@@ -724,9 +736,13 @@ class LiveSpeakerIdentifier {
};
}
- speakerId = this.currentSegmentSpeakerId || this._assignOrForceCluster(embedding);
- if (updateCentroid && this.currentSegmentSpeakerId) {
- this._updateCentroid(speakerId, embedding);
+ speakerId = this.currentSegmentSpeakerId;
+ if (speakerId) {
+ if (updateCentroid) {
+ this._updateCentroid(speakerId, embedding);
+ }
+ } else {
+ ({ speakerId } = this._assignOrForceCluster(embedding));
}
return {
@@ -743,10 +759,10 @@ class LiveSpeakerIdentifier {
// overflow is usually a different voice, and a polluted centroid would
// keep capturing it even after the cap rises (participants added
// mid-meeting).
- return nearest;
+ return { speakerId: nearest, forced: true };
}
}
- return this._assignSpeakerId(embedding);
+ return { speakerId: this._assignSpeakerId(embedding), forced: false };
}
_findNearestTransient(embedding) {
diff --git a/test/helpers/liveSpeakerIdentifier.test.js b/test/helpers/liveSpeakerIdentifier.test.js
index 4a0b49c6b..58822eb44 100644
--- a/test/helpers/liveSpeakerIdentifier.test.js
+++ b/test/helpers/liveSpeakerIdentifier.test.js
@@ -204,6 +204,35 @@ test("a voice force-merged at the cap becomes its own speaker once the cap rises
);
});
+test("an at-cap fold never steals the cluster's identity for a profile-matched voice", () => {
+ const { LiveSpeakerIdentifier } = loadIdentifier();
+ const identifier = new LiveSpeakerIdentifier();
+ identifier.setMaxSpeakers(1);
+ identifier.getSpeakerProfiles = () => [
+ { id: 1, display_name: "Alice", embedding: voiceA },
+ { id: 2, display_name: "Bob", embedding: voiceB },
+ ];
+
+ const alice = identifier._resolveSpeakerForEmbedding(voiceA, { updateCentroid: true });
+ assert.equal(alice.displayName, "Alice");
+
+ // Bob overflows the cap: his utterance is folded into Alice's cluster, but
+ // the cluster must keep Alice's identity — retagging it as Bob would both
+ // mislabel Alice and re-capture Bob through the profile lookup later.
+ const bobAtCap = identifier._resolveSpeakerForEmbedding(voiceB, { updateCentroid: true });
+ assert.equal(bobAtCap.speakerId, alice.speakerId);
+ assert.equal(bobAtCap.displayName, "Alice");
+
+ identifier.setMaxSpeakers(3);
+ const bobAfter = identifier._resolveSpeakerForEmbedding(voiceB, { updateCentroid: true });
+ assert.notEqual(bobAfter.speakerId, alice.speakerId, "Bob must get his own cluster");
+ assert.equal(bobAfter.displayName, "Bob");
+
+ const aliceAgain = identifier._resolveSpeakerForEmbedding(voiceA, { updateCentroid: true });
+ assert.equal(aliceAgain.speakerId, alice.speakerId);
+ assert.equal(aliceAgain.displayName, "Alice", "Alice's cluster must still be hers");
+});
+
test("distinct voices are folded into one cluster when the session cap is 1", () => {
const { LiveSpeakerIdentifier } = loadIdentifier();
const identifier = new LiveSpeakerIdentifier();