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 MAX_CANDIDATES = 5;
// Case, accents, spaces and punctuation don't tell apart the names people
// say aloud: "Eng Backend", "eng-backend" and "eng_backend" are one name.
function normalizeName(value) {
return String(value ?? "")
.normalize("NFKD")
.replace(/\p{M}/gu, "")
.toLowerCase()
.replace(/[^\p{L}\p{N}]/gu, "");
}
function distinctById(candidates) {
const byId = new Map();
for (const candidate of candidates) {
if (!byId.has(candidate.id)) byId.set(candidate.id, candidate);
}
return [...byId.values()];
}
// Stages run in order and the first stage with any match decides: an exact
// name never loses to a longer one that merely starts the same way, and two
// matches at the same stage are always ambiguous, never guessed. exactOnly
// is for incomplete candidate lists, where a looser match elsewhere could
// have been missed.
function resolveTarget(query, candidates, { exactOnly = false } = {}) {
const raw = String(query ?? "").trim();
const normalized = normalizeName(raw);
if (!normalized) return { status: "none" };
const stages = [
(name) => name === raw,
(name) => normalizeName(name) === normalized,
(name) => normalizeName(name).startsWith(normalized),
];
for (const matches of exactOnly ? stages.slice(0, 1) : stages) {
const found = distinctById(
candidates.filter((candidate) =>
candidate.names.some((name) => Boolean(name) && matches(name))
)
);
if (found.length === 1) return { status: "match", candidate: found[0] };
if (found.length > 1)
return { status: "ambiguous", candidates: found.slice(0, MAX_CANDIDATES) };
}
return { status: "none" };
}
module.exports = { resolveTarget, normalizeName, MAX_CANDIDATES };