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
import type { Segment, Speaker } from '@/data/types';
import { getSpeakerDisplayName } from '@/lib/diarization/transcriptDisplay';
import { hasFiniteNonZeroNorm } from '@/lib/diarization/voiceprints';
import { ENROLLMENT_MIN_SPEECH_ACTIVITY_MS } from '@/services/diarization/VoiceprintService';
export interface VoiceSetupCandidate {
speakerId: number;
name: string;
speechMs: number;
sampleLine: string;
}
/**
* The meeting-note banner. It only shows when That's me can work: the meeting was diarized
* on this device and a speaker's sample is still held from it (samples don't survive a
* restart), so it never opens onto an empty sheet.
*/
export function shouldOfferVoiceSetup({
isOnDeviceMeeting,
transcriptStatus,
hasOwnerProfile,
dismissed,
candidateCount,
}: {
isOnDeviceMeeting: boolean;
transcriptStatus: string;
hasOwnerProfile: boolean;
dismissed: boolean;
candidateCount: number;
}): boolean {
return (
isOnDeviceMeeting &&
transcriptStatus === 'done' &&
!hasOwnerProfile &&
!dismissed &&
candidateCount > 0
);
}
/**
* Speakers who said enough, with a sample still held from this meeting, longest first. A
* speaker still pointing at a deleted profile counts as unlinked, so you can claim it again.
*/
export function voiceSetupCandidates({
segments,
speakers,
embeddingsByLabel,
profileIds,
}: {
segments: Segment[];
speakers: Speaker[];
embeddingsByLabel: Record<string, number[]> | undefined;
profileIds: ReadonlySet<number>;
}): VoiceSetupCandidate[] {
return speakers
.filter(
(speaker) =>
(speaker.profileId == null || !profileIds.has(speaker.profileId)) &&
hasFiniteNonZeroNorm(embeddingsByLabel?.[speaker.speakerLabel]),
)
.map((speaker) => {
const own = segments.filter((segment) => segment.speakerLabel === speaker.speakerLabel);
const speechMs = own.reduce(
(total, segment) => total + Math.max(0, segment.endMs - segment.startMs),
0,
);
const sampleLine = own
.map((segment) => segment.text.trim())
.reduce((longest, text) => (text.length > longest.length ? text : longest), '');
return {
speakerId: speaker.id,
name: getSpeakerDisplayName(speaker, speaker.sortOrder),
speechMs,
sampleLine,
};
})
.filter((candidate) => candidate.speechMs >= ENROLLMENT_MIN_SPEECH_ACTIVITY_MS)
.sort((a, b) => b.speechMs - a.speechMs);
}
/** "45 sec", "3 min": how long someone spoke, not a position in the recording. */
export function formatSpeakingTime(ms: number): string {
const seconds = Math.round(ms / 1000);
return seconds < 60 ? `${seconds} sec` : `${Math.round(seconds / 60)} min`;
}