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
// PCM16 mono at the native tap's 24kHz. Reuse bounded silence blocks even if
// recovery takes a long time; never allocate the entire missing interval.
const SAMPLE_RATE = 24000;
// Match normal 100ms tap delivery. Providers may prepend a pending partial
// packet, so a block at their maximum frame size would overflow that limit.
const SILENCE_SAMPLES = SAMPLE_RATE / 10;
const SILENCE = Buffer.alloc(SILENCE_SAMPLES * 2);
const createMeetingAudioTimeline = ({ now = () => performance.now(), wallNow = Date.now } = {}) => {
let startedAt = null;
let wallStartedAt = null;
let samples = 0;
let recovering = false;
return {
markRestart() {
recovering = true;
},
write(buffer, emit) {
const receivedAt = now();
startedAt ??= receivedAt;
wallStartedAt ??= wallNow();
if (recovering) {
// Repair only explicit capture restarts. Arrival jitter and buffered
// delivery during normal capture must not manufacture gaps.
const elapsedSamples = Math.round(((receivedAt - startedAt) * SAMPLE_RATE) / 1000);
let missingSamples = Math.max(0, elapsedSamples - samples);
while (missingSamples > 0) {
const count = Math.min(missingSamples, SILENCE_SAMPLES);
emit(
SILENCE.subarray(0, count * 2),
true,
wallStartedAt + (samples * 1000) / SAMPLE_RATE
);
samples += count;
missingSamples -= count;
}
recovering = false;
}
const capturedAt = wallStartedAt + (samples * 1000) / SAMPLE_RATE;
samples += buffer.length / 2;
emit(buffer, false, capturedAt);
},
};
};
module.exports = createMeetingAudioTimeline;