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");
const {
timestampRequestFields,
mapVerboseSegments,
} = require("../../src/helpers/uploadTimestamps");
test("timestamp fields are requested only from providers/models that accept them", () => {
assert.deepEqual(timestampRequestFields("openai", "whisper-1"), {
response_format: "verbose_json",
});
assert.deepEqual(timestampRequestFields("groq", "whisper-large-v3"), {
response_format: "verbose_json",
});
assert.deepEqual(timestampRequestFields("groq", "whisper-large-v3-turbo"), {
response_format: "verbose_json",
});
assert.deepEqual(timestampRequestFields("mistral", "voxtral-mini-latest"), {
timestamp_granularities: "segment",
});
// gpt-4o transcribe models reject verbose_json outright.
assert.equal(timestampRequestFields("openai", "gpt-4o-transcribe"), null);
assert.equal(timestampRequestFields("openai", "gpt-4o-mini-transcribe"), null);
// Everything else fails safe to a plain-text request: a custom or
// self-hosted server may 400 on fields it doesn't know.
assert.equal(timestampRequestFields("custom", "whisper-1"), null);
assert.equal(timestampRequestFields("self-hosted", "whisper-large-v3"), null);
assert.equal(timestampRequestFields("xai", "grok-stt"), null);
assert.equal(timestampRequestFields(undefined, undefined), null);
});
test("verbose segments map to {text, start, end}", () => {
assert.deepEqual(
mapVerboseSegments({
text: "Hello world. Goodbye.",
segments: [
{ id: 0, seek: 0, start: 0, end: 4.2, text: " Hello world." },
{ id: 1, seek: 400, start: 4.2, end: 7.5, text: " Goodbye." },
],
}),
[
{ text: "Hello world.", start: 0, end: 4.2 },
{ text: "Goodbye.", start: 4.2, end: 7.5 },
]
);
});
test("malformed verbose segments are dropped, never thrown", () => {
const mapped = mapVerboseSegments({
segments: [
{ start: 0, end: 1, text: " " },
{ start: NaN, end: 2, text: "no start" },
{ start: "1", end: 2, text: "string start" },
{ end: 2, text: "missing start" },
null,
{ start: 3, end: "x", text: "bad end falls back to start" },
{ start: 5, text: "missing end falls back to start" },
],
});
assert.deepEqual(mapped, [
{ text: "bad end falls back to start", start: 3, end: 3 },
{ text: "missing end falls back to start", start: 5, end: 5 },
]);
});
test("verbose segments clamp negative start timestamps and inverted end timestamps", () => {
const mapped = mapVerboseSegments({
segments: [
{ start: -2.5, end: 3, text: "negative start clamped" },
{ start: 6, end: 2, text: "inverted end clamped" },
],
});
assert.deepEqual(mapped, [
{ text: "negative start clamped", start: 0, end: 3 },
{ text: "inverted end clamped", start: 6, end: 6 },
]);
});
// The segments become the upload note's only copy of the text, so they must
// hold every word of it; a CJK response's text has no spaces between them.
test("segments must cover the response text, whitespace aside", () => {
assert.deepEqual(
mapVerboseSegments({
text: "你好。今天开会。",
segments: [
{ start: 0, end: 1, text: "你好。" },
{ start: 1, end: 2, text: "今天开会。" },
],
}),
[
{ text: "你好。", start: 0, end: 1 },
{ text: "今天开会。", start: 1, end: 2 },
]
);
assert.equal(
mapVerboseSegments({
text: "First sentence. Second sentence.",
segments: [
{ start: 0, end: 1, text: " First sentence." },
{ start: NaN, end: 2, text: " Second sentence." },
],
}),
null
);
});
test("responses without usable segments map to null", () => {
assert.equal(mapVerboseSegments({ text: "plain json response" }), null);
assert.equal(mapVerboseSegments({ segments: "not-an-array" }), null);
assert.equal(mapVerboseSegments({ segments: [] }), null);
assert.equal(mapVerboseSegments({ segments: [{ start: NaN, text: "" }] }), null);
assert.equal(mapVerboseSegments(null), null);
assert.equal(mapVerboseSegments(undefined), null);
});