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
const test = require("node:test");
const assert = require("node:assert/strict");
// The .ts import is stripped by tsx, which `npm test` loads via --import.
const load = () => import("../../src/utils/audioContainer.ts");
const readAscii = (view, offset, length) =>
Array.from({ length }, (_, i) => String.fromCharCode(view.getUint8(offset + i))).join("");
test("only the custom provider re-encodes, and only for containers a WAV/MP3/FLAC backend refuses", async () => {
const { needsWavConversion } = await load();
assert.equal(needsWavConversion("custom", "audio/webm;codecs=opus"), true);
assert.equal(needsWavConversion("custom", "audio/ogg"), true);
// Allowlisted rather than blocklisted: MediaRecorder emits MP4/AAC on some
// platforms, and AAC is no more acceptable to these backends than Opus.
assert.equal(needsWavConversion("custom", "audio/mp4"), true);
assert.equal(needsWavConversion("custom", "video/webm"), true);
// Already acceptable — re-encoding would only cost bytes.
assert.equal(needsWavConversion("custom", "audio/wav"), false);
assert.equal(needsWavConversion("custom", "audio/x-wav"), false);
assert.equal(needsWavConversion("custom", "audio/mpeg"), false);
assert.equal(needsWavConversion("custom", "audio/flac"), false);
// Built-in providers accept Opus today; leaving them alone keeps uploads small.
for (const provider of ["openai", "groq", "mistral", "xai", "gemini", "corti", "tinfoil"]) {
assert.equal(needsWavConversion(provider, "audio/webm"), false);
}
// Missing type must not trigger a pointless decode.
assert.equal(needsWavConversion("custom", undefined), false);
assert.equal(needsWavConversion("custom", ""), false);
// An empty recording has nothing to decode; attempting it would only log a
// misleading "re-encode failed" warning.
assert.equal(needsWavConversion("custom", "audio/webm", 0), false);
assert.equal(needsWavConversion("custom", "audio/webm", 1024), true);
});
test("downmixToMono averages channels and passes mono through untouched", async () => {
const { downmixToMono } = await load();
const mono = new Float32Array([0.25, -0.5]);
assert.equal(downmixToMono([mono]), mono, "mono input is returned as-is, not copied");
const mixed = downmixToMono([new Float32Array([1, -1]), new Float32Array([0, 1])]);
assert.deepEqual(Array.from(mixed), [0.5, 0]);
assert.throws(() => downmixToMono([]), /at least one channel/);
});
test("encodeWav writes a RIFF header the upstream decoder can read", async () => {
const { encodeWav } = await load();
const frames = 8;
const buffer = encodeWav([new Float32Array(frames)], 48000);
const view = new DataView(buffer);
assert.equal(readAscii(view, 0, 4), "RIFF");
assert.equal(readAscii(view, 8, 4), "WAVE");
assert.equal(readAscii(view, 12, 4), "fmt ");
assert.equal(readAscii(view, 36, 4), "data");
assert.equal(view.getUint16(20, true), 1, "PCM format tag");
assert.equal(view.getUint16(22, true), 1, "channel count");
assert.equal(view.getUint32(24, true), 48000, "sample rate");
assert.equal(view.getUint32(28, true), 48000 * 2, "byte rate");
assert.equal(view.getUint16(32, true), 2, "block align");
assert.equal(view.getUint16(34, true), 16, "bits per sample");
const dataBytes = frames * 2;
assert.equal(view.getUint32(40, true), dataBytes, "data chunk size");
assert.equal(view.getUint32(4, true), 36 + dataBytes, "riff size");
assert.equal(buffer.byteLength, 44 + dataBytes);
});
test("samples outside [-1, 1] clamp to the rails instead of wrapping", async () => {
const { encodeWav } = await load();
// decodeAudioData can return values slightly past full scale; a naive
// multiply-and-truncate turns +1.2 into a large negative sample.
const view = new DataView(encodeWav([new Float32Array([1.2, -1.2, 1, -1, 0])], 16000));
assert.equal(view.getInt16(44, true), 32767);
assert.equal(view.getInt16(46, true), -32768);
assert.equal(view.getInt16(48, true), 32767);
assert.equal(view.getInt16(50, true), -32768);
assert.equal(view.getInt16(52, true), 0);
});
test("multi-channel audio is interleaved per frame", async () => {
const { encodeWav } = await load();
const left = new Float32Array([1, 0]);
const right = new Float32Array([0, -1]);
const view = new DataView(encodeWav([left, right], 44100));
assert.equal(view.getUint16(22, true), 2, "channel count");
assert.equal(view.getUint32(28, true), 44100 * 2 * 2, "byte rate");
assert.equal(view.getInt16(44, true), 32767, "frame 0 left");
assert.equal(view.getInt16(46, true), 0, "frame 0 right");
assert.equal(view.getInt16(48, true), 0, "frame 1 left");
assert.equal(view.getInt16(50, true), -32768, "frame 1 right");
});
test("encodeWav refuses an empty channel list rather than emitting a headerless file", async () => {
const { encodeWav } = await load();
assert.throws(() => encodeWav([], 16000), /at least one channel/);
});