forked from Blendman974/faceclaw
Add symbolic math and graphing tools, dual-microphone audio processing, live captions, translation, speaker diarization, sentiment analysis, and searchable conversation history. Extend Android voice and BLE support for microphone control, recording, transcoding, and on-device speaker identification, with coverage for calculator and microphone workflows.
80 lines
2.9 KiB
TypeScript
80 lines
2.9 KiB
TypeScript
import { isLocalModelReady, streamLocalQwen } from "../../native/llama";
|
|
import type { NameInference } from "./name-guess";
|
|
|
|
/**
|
|
* Second-tier verification of heuristic name detections using the on-phone
|
|
* LLM (the assistant's Qwen model, when downloaded). The regex tier is
|
|
* deliberately permissive about unusual real names; the model is asked the
|
|
* one question the regex can't answer — "is this actually a person's name
|
|
* being introduced here?" — before a speaker profile gets renamed.
|
|
*
|
|
* Verdicts: true (confirmed), false (rejected), null (model unavailable,
|
|
* busy, timed out, or gave a non-answer). Callers treat null as "trust the
|
|
* heuristic": renames are retroactively correctable, so a missing model
|
|
* must not disable introduction naming entirely.
|
|
*/
|
|
|
|
const VERDICT_TIMEOUT_MS = 25_000;
|
|
|
|
// One verification at a time: the runner is shared with the assistant, and
|
|
// caption lines can arrive faster than CPU inference finishes.
|
|
let inFlight = false;
|
|
|
|
export function verifyIntroducedName(
|
|
inference: NameInference,
|
|
sentence: string,
|
|
): Promise<boolean | null> {
|
|
if (!global.isAndroid || !isLocalModelReady() || inFlight) {
|
|
return Promise.resolve(null);
|
|
}
|
|
inFlight = true;
|
|
return new Promise((resolve) => {
|
|
let settled = false;
|
|
const finish = (verdict: boolean | null) => {
|
|
if (settled) return;
|
|
settled = true;
|
|
inFlight = false;
|
|
resolve(verdict);
|
|
};
|
|
const timer = setTimeout(() => finish(null), VERDICT_TIMEOUT_MS);
|
|
try {
|
|
streamLocalQwen({
|
|
apiKey: "",
|
|
model: "local",
|
|
system:
|
|
"You verify speaker introductions for a live-captioning app. " +
|
|
"Given a transcript line and a candidate name, decide whether the candidate " +
|
|
"is genuinely a person's name being introduced (their own or someone else's), " +
|
|
"rather than an ordinary word, feeling, or phrase. " +
|
|
'Answer with exactly one word: "YES" or "NO".',
|
|
messages: [
|
|
{
|
|
role: "user",
|
|
content:
|
|
`Transcript line: "${sentence.trim()}"\n` +
|
|
`Candidate name: "${inference.name}" ` +
|
|
`(detected as a ${inference.kind === "self" ? "self-introduction" : "third-party introduction"}).\n` +
|
|
`Is "${inference.name}" a person's name being introduced in this line?`,
|
|
},
|
|
],
|
|
maxTokens: 8,
|
|
onDone: (result) => {
|
|
clearTimeout(timer);
|
|
const text = result.text.trim().toUpperCase();
|
|
if (/\bYES\b/.test(text)) finish(true);
|
|
else if (/\bNO\b/.test(text)) finish(false);
|
|
else finish(null);
|
|
},
|
|
onError: () => {
|
|
clearTimeout(timer);
|
|
finish(null);
|
|
},
|
|
});
|
|
} catch (error) {
|
|
clearTimeout(timer);
|
|
console.warn(`name verification failed: ${error}`);
|
|
finish(null);
|
|
}
|
|
});
|
|
}
|