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 { TranscriptSegment } from "../stores/meetingRecordingStore";
import type { CalendarAttendee } from "../types/calendar";
import { isYouSpeakerName, resolveSegmentSpeakerName } from "./transcriptSpeakerState";
import type { MentionPerson } from "./mentionMarkdown";
type Translate = (key: string, options?: Record<string, unknown>) => string;
export interface MeetingIdentity {
selfName: string | null;
selfEmail: string | null;
participants: CalendarAttendee[];
}
const isSelfParticipant = (participant: CalendarAttendee, identity: MeetingIdentity) =>
participant.self || (!!identity.selfEmail && participant.email === identity.selfEmail);
/**
* Speaker label for the LLM payload. An unnamed, unnumbered mic track falls back
* to the note owner's real name instead of an anonymous "You" whenever it is known.
* Diarization numbers the mic's clusters when several people share it (an
* in-person meeting), so a numbered mic segment is someone in the room, not
* necessarily the owner — matching the "Speaker N" the transcript view shows.
*/
export function resolveLlmSpeakerLabel(
segment: TranscriptSegment,
speakerMappings: Record<string, string>,
selfLabel: string,
t: Translate
): string {
const resolved = resolveSegmentSpeakerName(segment, speakerMappings);
if (resolved) return isYouSpeakerName(resolved, t("notes.speaker.you")) ? selfLabel : resolved;
if (segment.speaker === "you") return selfLabel;
if (segment.speaker) {
const num = Number.parseInt(segment.speaker.replace("speaker_", ""), 10);
if (!Number.isNaN(num)) return t("notes.speaker.label", { n: num + 1 });
}
if (segment.source === "mic") return selfLabel;
return t("notes.speaker.them");
}
export function buildLlmTranscript(
segments: TranscriptSegment[],
speakerMappings: Record<string, string>,
selfLabel: string,
t: Translate
): string {
return segments
.map((s) => `${resolveLlmSpeakerLabel(s, speakerMappings, selfLabel, t)}: ${s.text}`)
.join("\n");
}
const formatPerson = (name: string | null, email: string | null): string => {
const trimmedName = name?.trim() || "";
const trimmedEmail = email?.trim() || "";
if (trimmedName && trimmedEmail) return `${trimmedName} <${trimmedEmail}>`;
return trimmedName || trimmedEmail;
};
/**
* Context block prepended to the LLM payload so note generation knows who the
* note owner is (instead of guessing an identity from names spoken in the
* transcript) and who was invited.
*/
export function buildMeetingContext(identity: MeetingIdentity, selfLabel: string): string {
const lines: string[] = [];
const self = formatPerson(identity.selfName, identity.selfEmail);
if (self) {
lines.push(`The user taking these notes ("${selfLabel}" in the transcript) is ${self}.`);
}
const invited = identity.participants
.filter((p) => !isSelfParticipant(p, identity))
.map((p) => formatPerson(p.displayName, p.email))
.filter(Boolean);
if (invited.length > 0) lines.push(`Invited participants: ${invited.join(", ")}.`);
return lines.length > 0 ? `## Meeting Context\n${lines.join("\n")}` : "";
}
/**
* Everyone we can name around a note — the note owner, calendar participants,
* mapped speaker clusters, and named segments. Same-name attendees with
* different emails stay distinct so tagging can detect ambiguous owners.
* Feeds the editor's @mention suggestions and action-item owner tagging.
*/
export function collectKnownPeople(
identity: MeetingIdentity,
speakerMappings: Record<string, string>,
segments: TranscriptSegment[]
): MentionPerson[] {
const people: MentionPerson[] = [];
const byName = new Map<string, MentionPerson[]>();
const push = (name: string | null | undefined, email: string | null | undefined) => {
const trimmed = name?.trim();
if (!trimmed) return;
const key = trimmed.replace(/\s+/g, " ").toLowerCase();
const address = email?.trim() || null;
const existing = byName.get(key) ?? [];
if (existing.length > 0) {
if (!address || existing.some((p) => p.email?.toLowerCase() === address.toLowerCase()))
return;
}
const person = { name: trimmed, email: address };
existing.push(person);
byName.set(key, existing);
people.push(person);
};
push(identity.selfName, identity.selfEmail);
for (const p of identity.participants) {
if (isSelfParticipant(p, identity)) continue;
push(p.displayName || p.email.split("@")[0], p.email);
}
for (const name of Object.values(speakerMappings)) push(name, null);
for (const segment of segments) push(segment.speakerName, null);
return people;
}