Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.
JavaScript aianthropiccross-platformgeminigroqlinuxmacosnvidiaopen-sourceopenaiparakeetspeech-to-texttranscribewhisperwindows
fix(meeting): anchor live speaker timestamps to the first system chunk
Gabriel Steina221ea46acaf
diff --git a/src/helpers/ipcHandlers.js b/src/helpers/ipcHandlers.jsindex 24e6d8e41..0c4292df6 100644--- a/src/helpers/ipcHandlers.js+++ b/src/helpers/ipcHandlers.js@@ -6180,58 +6180,71 @@ class IPCHandlers {}meetingLiveSpeakerState = null;- meetingLiveSpeakerStartedAt = Date.now();- const started = await liveSpeakerIdentifier.start(- (identification) => {- if (!win || win.isDestroyed()) {- return;- }+ // Anchored on the first system chunk instead, in sendMeetingAudio.+ meetingLiveSpeakerStartedAt = null;+ const started = await liveSpeakerIdentifier+ .start(+ (identification) => {+ if (!win || win.isDestroyed() || meetingLiveSpeakerStartedAt == null) {+ return;+ }- bindOneOnOneAttendeeToSpeaker(identification.speakerId);+ bindOneOnOneAttendeeToSpeaker(identification.speakerId);- const displayName = meetingOneOnOneAttendee- ? meetingOneOnOneAttendee.displayName- : identification.displayName;+ const displayName = meetingOneOnOneAttendee+ ? meetingOneOnOneAttendee.displayName+ : identification.displayName;- const startTime = Math.max(- meetingLiveSpeakerStartedAt || 0,- (meetingLiveSpeakerStartedAt || 0) + identification.startTime * 1000- );- const endTime = Math.max(- startTime,- (meetingLiveSpeakerStartedAt || 0) + identification.endTime * 1000- );- const enrichedIdentification = {- ...identification,- displayName,- startTime,- endTime,- };+ const startTime = Math.max(+ meetingLiveSpeakerStartedAt,+ meetingLiveSpeakerStartedAt + identification.startTime * 1000+ );+ const endTime = Math.max(+ startTime,+ meetingLiveSpeakerStartedAt + identification.endTime * 1000+ );+ const enrichedIdentification = {+ ...identification,+ displayName,+ startTime,+ endTime,+ };- win.webContents.send("meeting-speaker-identified", enrichedIdentification);-- for (const seg of meetingDiarizationSegments) {- if (- seg.source === "system" &&- seg.timestamp != null &&- seg.timestamp >= startTime &&- seg.timestamp <= endTime &&- (!seg.speaker || seg.speakerIsPlaceholder)- ) {- applyConfirmedSpeaker(seg, {- speaker: identification.speakerId,- speakerName: displayName || seg.speakerName,- speakerIsPlaceholder: false,- });+ win.webContents.send("meeting-speaker-identified", enrichedIdentification);++ for (const seg of meetingDiarizationSegments) {+ if (+ seg.source === "system" &&+ seg.timestamp != null &&+ seg.timestamp >= startTime &&+ seg.timestamp <= endTime &&+ (!seg.speaker || seg.speakerIsPlaceholder)+ ) {+ applyConfirmedSpeaker(seg, {+ speaker: identification.speakerId,+ speakerName: displayName || seg.speakerName,+ speakerIsPlaceholder: false,+ });+ }}+ },+ {+ getSpeakerProfiles: getLiveSpeakerProfiles,+ maxSpeakers: resolveSessionMaxSpeakers(),+ enabled: true,}- },- {- getSpeakerProfiles: getLiveSpeakerProfiles,- maxSpeakers: resolveSessionMaxSpeakers(),- enabled: true,- }- );+ )+ .catch((error) => {+ // isAvailable() only stats the model file, so a corrupt model or an+ // onnxruntime binding that won't load still throws here. Speaker labels+ // are an enhancement — never let them take the recording down with them.+ debugLogger.warn(+ "Live speaker identification start failed",+ { error: error.message },+ "speaker"+ );+ return false;+ });if (started) {meetingLiveSpeakerActive = true;@@ -6956,6 +6969,13 @@ class IPCHandlers {flushPendingMeetingMicChunks();if (meetingLiveSpeakerActive) {+ // identification.startTime counts samples from the first chunk the+ // identifier sees, so the wall-clock anchor has to be the arrival of+ // that chunk. Stamping it when identification starts is only correct+ // when capture is already running (the macOS tap); the Windows+ // loopback helper can take seconds to hand over its first buffer, and+ // a stale anchor shifts every label earlier by that gap.+ meetingLiveSpeakerStartedAt ??= receivedAt;void liveSpeakerIdentifier.feedAudio(outboundBuffer);}diff --git a/test/helpers/liveSpeakerIdPolicy.test.js b/test/helpers/liveSpeakerIdPolicy.test.jsindex d19ed72fb..58903b193 100644--- a/test/helpers/liveSpeakerIdPolicy.test.js+++ b/test/helpers/liveSpeakerIdPolicy.test.js@@ -41,11 +41,32 @@ test("missing mode never identifies speakers", async () => {assert.equal(supportsLiveSpeakerIdentification(null, "darwin"), false);});+// The sole production call site passes only the mode, so the defaulted platform+// is the path that actually ships. Asserting it against process.platform would+// compare the function to itself and pass on CI (Linux) even if the default were+// deleted, so stub the platform instead.test("platform defaults to the current process platform", async () => {const { supportsLiveSpeakerIdentification } = await load();+ const originalPlatform = Object.getOwnPropertyDescriptor(process, "platform");- assert.equal(- supportsLiveSpeakerIdentification("loopback"),- supportsLiveSpeakerIdentification("loopback", process.platform)- );+ const onPlatform = (platform, run) => {+ Object.defineProperty(process, "platform", { value: platform, configurable: true });+ try {+ run();+ } finally {+ Object.defineProperty(process, "platform", originalPlatform);+ }+ };++ onPlatform("win32", () => {+ assert.equal(supportsLiveSpeakerIdentification("loopback"), true);+ assert.equal(supportsLiveSpeakerIdentification("native"), true);+ });+ onPlatform("linux", () => {+ assert.equal(supportsLiveSpeakerIdentification("loopback"), false);+ assert.equal(supportsLiveSpeakerIdentification("native"), true);+ });+ onPlatform("darwin", () => {+ assert.equal(supportsLiveSpeakerIdentification("loopback"), false);+ });});