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
export interface Snippet {
trigger: string;
replacement: string;
}
// Mirrors the cap database.js enforces when it stores snippets.
export const MAX_SNIPPET_TRIGGER_LENGTH = 100;
interface SnippetMatcher {
regex: RegExp;
replacements: Map<string, string>;
}
export interface SnippetTriggerRange {
/** Index of the trigger's first character. */
start: number;
/** Index one past the trigger's last character. */
end: number;
}
let cachedSnippets: Snippet[] | null = null;
let cachedMatcher: SnippetMatcher | null = null;
// The regex /i flag can't case-fold Turkish İ (U+0130) or dotless ı
// (U+0131), and İ's toLowerCase() form is two code units ("i" + U+0307), so
// triggers like "İmza" never matched. Folding İ to a plain "i" gives Map
// keys a canonical form, and matching İ/ı explicitly in the pattern lets the
// regex find them in the transcript.
function foldCapitalIDot(value: string): string {
return value.replace(/İ/g, "i");
}
function buildMatcher(snippets: Snippet[]): SnippetMatcher | null {
const replacements = new Map<string, string>();
for (const { trigger, replacement } of snippets) {
const folded = foldCapitalIDot(trigger.trim().normalize("NFC"));
const key = folded.toLowerCase();
if (!key) continue;
replacements.set(key, replacement);
// An uppercase I in the trigger may mean Turkish ı as well as English i,
// so register both readings; an explicit trigger wins over a variant.
const dotlessKey = folded.replace(/I/g, "ı").toLowerCase();
if (!replacements.has(dotlessKey)) replacements.set(dotlessKey, replacement);
}
if (replacements.size === 0) return null;
// Longest-first so "investor ask" wins over a shorter "ask" trigger.
const escaped = [...replacements.keys()]
.sort((a, b) => b.length - a.length)
.map((t) => t.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"))
.map((t) => t.replace(/i/g, "[iİ]").replace(/ı/g, "[ıI]"));
// Unicode-aware word boundaries — triggers never match inside a word.
const regex = new RegExp(
`(?<=^|[\\s\\p{P}\\p{S}])(?:${escaped.join("|")})(?=$|[\\s\\p{P}\\p{S}])`,
"giu"
);
return { regex, replacements };
}
// Memoized against the snippets array reference (the settings store replaces
// the array on every change).
function getMatcher(snippets?: Snippet[] | null): SnippetMatcher | null {
if (!Array.isArray(snippets) || snippets.length === 0) return null;
if (snippets !== cachedSnippets) {
cachedSnippets = snippets;
cachedMatcher = buildMatcher(snippets);
}
return cachedMatcher;
}
// The pattern case-folds the Unicode way (/iu) while the keys are toLowerCase'd,
// so a match is not proof of a replacement: "[ıI]mza" matches a plain "imza",
// which is no key at all. Both callers resolve through here, so a reported range
// always means an expansion — wake-word suppression depends on it.
function resolveReplacement(match: string, replacements: Map<string, string>): string | undefined {
const folded = foldCapitalIDot(match);
return (
replacements.get(folded.toLowerCase()) ??
// An uppercase I can be capital dotless ı as well as English i.
replacements.get(folded.replace(/I/g, "ı").toLowerCase())
);
}
/**
* Character ranges of every trigger occurrence, matched against `text` exactly
* as given so the offsets index that same string. Wake-word detection uses
* these to ignore an agent name that only appears because it opens a trigger
* the user chose, such as "openwhispr review" (see `agentDetection`).
*/
export function findSnippetTriggerRanges(
text: string,
snippets?: Snippet[] | null
): SnippetTriggerRange[] {
if (!text) return [];
const matcher = getMatcher(snippets);
if (!matcher) return [];
return [...text.matchAll(matcher.regex)]
.filter((match) => resolveReplacement(match[0], matcher.replacements) !== undefined)
.map((match) => ({
start: match.index,
end: match.index + match[0].length,
}));
}
/**
* Replace every spoken trigger with its saved text in a single pass.
*/
export function expandSnippets(text: string, snippets?: Snippet[] | null): string {
if (!text) return text;
const matcher = getMatcher(snippets);
if (!matcher) return text;
const { regex, replacements } = matcher;
// NFC so a decomposed "I" + U+0307 in the transcript recombines into İ.
return text
.normalize("NFC")
.replace(regex, (match) => resolveReplacement(match, replacements) ?? match);
}
/**
* Dictionary words plus snippet triggers — the hint list fed to the STT
* prompt and cleanup-model dictionary suffix so triggers survive both.
*/
export function getDictionaryHintWords(
settings?: {
customDictionary?: string[] | null;
snippets?: Snippet[] | null;
} | null
): string[] {
const dictionary = Array.isArray(settings?.customDictionary) ? settings.customDictionary : [];
const snippets = Array.isArray(settings?.snippets) ? settings.snippets : [];
if (snippets.length === 0) return [...dictionary];
const triggers = snippets
.map((s) => s?.trigger)
.filter((t): t is string => typeof t === "string" && t.length > 0);
return [...dictionary, ...triggers];
}