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
import { AppleLLM } from '@/lib/appleLLM';
import {
LOCAL_OUTPUT_TOKEN_RESERVE,
countLocalReasoningTokens,
estimateLocalTokens,
toLocalReasoningError,
} from '@/lib/localReasoning';
import type { ReasoningRequest, ReasoningResponse } from '@/types';
// Callers fold language, tone and dictionary into the system prompt first
// (buildProviderPrompt), so the instructions are that prompt alone.
export function buildLocalReasoningInstructions(request: ReasoningRequest): string {
return request.systemPrompt?.trim() ?? '';
}
export class LocalReasoningService {
static async processText(request: ReasoningRequest): Promise<ReasoningResponse> {
const instructions = buildLocalReasoningInstructions(request);
if (!instructions) {
throw new Error('Local reasoning requires explicit instructions.');
}
try {
const startedAt = Date.now();
const result = await AppleLLM.generateText(
{
instructions,
prompt: request.text,
temperature: request.temperature,
maxTokens: request.maxTokens ?? LOCAL_OUTPUT_TOKEN_RESERVE,
},
{ signal: request.signal },
);
if (__DEV__) {
// The rate the On-Device cleanup timeout assumes (cleanupTranscript); it includes reading
// the prompt, and is estimated from length where iOS can't count tokens (before 26.4).
const elapsedMs = Date.now() - startedAt;
const counted = await countLocalReasoningTokens({ prompt: result.text });
const outputTokens = counted ?? estimateLocalTokens(result.text);
console.log(
`[reasoning] provider=local model=apple-fm elapsedMs=${elapsedMs} outputTokens=${outputTokens}${
counted === null ? ' (estimated)' : ''
} tokensPerSecond=${(outputTokens / Math.max(elapsedMs / 1000, 0.001)).toFixed(1)}`,
);
}
return {
text: result.text.trim(),
model: 'apple-fm',
};
} catch (error) {
throw toLocalReasoningError(error);
}
}
}