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
// Map a reasoning cloud routing to the InferenceMode its Settings tab selects on.
// Mirrors deriveTranscriptionMode (byok custom → self-hosted, other cloud → providers).
export function deriveReasoningMode(cloudMode, provider) {
if (cloudMode === "byok") {
return provider === "custom" ? "self-hosted" : "providers";
}
return "openwhispr";
}
// Whether a scope may borrow the fallback scope's API key along with its endpoint.
// A scope pointing somewhere of its own, or in another mode, would send that key to
// a host it was never entered for.
export function inheritsFallbackEndpoint(own, fallbackMode) {
if (own.cloudBaseUrl || own.remoteUrl) return false;
return !!fallbackMode && own.mode === fallbackMode;
}
const MIRRORED_ROUTING_FIELDS = [
["cleanupProvider", "provider"],
["cleanupModel", "model"],
["cleanupCloudMode", "cloudMode"],
["cleanupCloudBaseUrl", "cloudBaseUrl"],
["cleanupRemoteUrl", "remoteUrl"],
["cleanupCustomApiKey", "customApiKey"],
];
// Fan a cleanup config out to all five LLM scopes; the four non-cleanup scopes
// mirror the routing fields that are set plus the derived mode (each tab selects
// on its mode). The endpoint and key ride along so a self-hosted or custom
// endpoint is reachable from every scope, not just the one onboarding wrote.
export function buildReasoningScopePatches(settings, mode) {
const dictationCleanup = { ...settings, cleanupMode: mode };
const routing = Object.fromEntries(
MIRRORED_ROUTING_FIELDS.filter(([cleanupKey]) => settings[cleanupKey] !== undefined).map(
([cleanupKey, field]) => [field, settings[cleanupKey]]
)
);
return {
dictationCleanup,
noteFormatting: { mode, ...routing },
dictationAgent: { mode, ...routing },
chatIntelligence: { mode, ...routing },
dictationTranslation: { mode, ...routing },
};
}
// Onboarding "use Corti everywhere" payloads. Transcription always routes to
// Corti. Reasoning routes to Corti only in the EU region with an API key, since
// Corti Models is EU-only and needs its own key; otherwise it routes to the
// HIPAA-compliant OpenWhispr Cloud so clinical text never reaches a third party.
// useCleanupModel is forced true either way so the routing sticks.
export function buildCortiOnboardingPayloads(
transcriptionProvider,
reasoningProvider,
environment,
hasApiKey
) {
const transcription = {
useLocalWhisper: false,
cloudTranscriptionMode: "byok",
cloudTranscriptionProvider: "corti",
cloudTranscriptionModel: transcriptionProvider?.models?.[0]?.id,
};
const cortiModel = reasoningProvider?.models?.[0]?.id;
const reasoning =
environment === "eu" && hasApiKey && cortiModel
? {
useCleanupModel: true,
cleanupProvider: "corti",
cleanupModel: cortiModel,
cleanupCloudMode: "byok",
}
: { useCleanupModel: true, cleanupCloudMode: "openwhispr" };
return { transcription, reasoning };
}