Claude
48fa6b4f9e
#74 (22 Teilnehmer) ergab automatisch 44 Sprecher, dieselbe Stimme mehrfach. Ursache gemessen: ein Zentroid pro sherpa-Cluster aus max. 3 Schnitten, fester Kosinus auf rohen Embeddings (Kreuz-Turn-Kosinus ~0,4 = keine Trennschaerfe). v2: sherpa liefert nur Turn-Grenzen, Identitaet aus 3-s-Fenster-Embeddings, Mittelwert-Zentrierung, Average-Linkage-AHC, Weak-Fold, Schwelle pro Aufnahme per 2-Gauss-EM gelernt. Validiert: #74 44->20, #70 ->4, #68 ->2. Live-Namensmarker: Namen waehrend der Aufnahme tippen; Zeitstempel benennt spaeter den Cluster und dient als must-/cannot-link-Zwangsbedingung.
72 lines
3.6 KiB
JavaScript
72 lines
3.6 KiB
JavaScript
// Speaker voiceprints: talks to the persistent spk_embed process (sherpa-onnx
|
|
// NeMo TitaNet embedding on CPU). embed(wav) → 192-dim vector; cosine match a
|
|
// diarization cluster against enrolled voiceprints → a name. All offline.
|
|
import { spawn } from 'node:child_process';
|
|
import fs from 'node:fs';
|
|
import readline from 'node:readline';
|
|
|
|
const BIN = '/opt/callrec/spk_embed';
|
|
const DIR = '/opt/sherpa/sherpa-onnx-v1.13.3-linux-aarch64-shared-cpu';
|
|
// Identity/clustering model. Benchmarked 2026-07-27 on our own conference audio
|
|
// (#74, 1181 mean-centred windows, same-turn vs cross-turn cosine):
|
|
// TitaNet-large d' 3.75 / 197 s CAM++ (3D-Speaker) d' 3.72 / 82 s
|
|
// TitaNet-small d' 3.54 / 57 s wespeaker VoxCeleb resnet34 d' 2.43 / 185 s
|
|
// Fed through the full v2 pipeline the three good models ended up within one
|
|
// speaker of each other (21/22/21 on a 23-person weekly), so TitaNet-small stays
|
|
// the default: it is the fastest AND keeps every already-learned voiceprint
|
|
// valid. SPK_MODEL switches models — prints are then stored per model (see
|
|
// db.upsertVoiceprint), because cosine across models is meaningless.
|
|
const EMB = process.env.SPK_MODEL || '/opt/sherpa/nemo_titanet_small.onnx';
|
|
export const THRESHOLD = +(process.env.SPK_THRESHOLD || 0.5);
|
|
// Voiceprints are only comparable within the SAME embedding model — stored with
|
|
// every print so a model swap can't silently match against stale vectors.
|
|
export function modelId() { return EMB.split('/').pop().replace(/\.onnx$/, ''); }
|
|
|
|
let proc = null, ready = null, rl = null, pending = null, chain = Promise.resolve();
|
|
|
|
function start() {
|
|
if (proc) return ready;
|
|
if (!fs.existsSync(EMB)) return Promise.reject(new Error('Embedding-Modell fehlt: ' + EMB));
|
|
proc = spawn('taskset', ['-c', '4-7', BIN, EMB],
|
|
{ env: { ...process.env, LD_LIBRARY_PATH: `${DIR}/lib` }, stdio: ['pipe', 'pipe', 'pipe'] });
|
|
proc.stderr.on('data', d => process.stderr.write('[spk] ' + d));
|
|
ready = new Promise((res, rej) => {
|
|
const to = setTimeout(() => rej(new Error('spk READY timeout')), 60000);
|
|
rl = readline.createInterface({ input: proc.stdout });
|
|
rl.on('line', line => {
|
|
if (line.startsWith('READY')) { clearTimeout(to); res(); return; }
|
|
if (!pending) return;
|
|
const p = pending; pending = null;
|
|
if (line.startsWith('OK ')) p.resolve(line.slice(3).split(',').map(Number));
|
|
else p.reject(new Error(line.replace(/^ERR\s*/, '') || 'spk error'));
|
|
});
|
|
});
|
|
const cleanup = () => { proc = null; ready = null; rl = null; if (pending) { pending.reject(new Error('spk beendet')); pending = null; } };
|
|
proc.on('exit', cleanup); proc.on('error', cleanup);
|
|
return ready;
|
|
}
|
|
|
|
// Extract the voiceprint for a 16 kHz mono wav file.
|
|
export function embed(wavPath) {
|
|
chain = chain.then(async () => {
|
|
await start();
|
|
return await new Promise((res, rej) => { pending = { resolve: res, reject: rej }; proc.stdin.write(wavPath + '\n'); });
|
|
});
|
|
return chain;
|
|
}
|
|
|
|
export function cosine(a, b) {
|
|
let d = 0, na = 0, nb = 0;
|
|
for (let i = 0; i < a.length; i++) { d += a[i] * b[i]; na += a[i] * a[i]; nb += b[i] * b[i]; }
|
|
return d / (Math.sqrt(na) * Math.sqrt(nb) + 1e-9);
|
|
}
|
|
|
|
// Best enrolled voiceprint for an embedding, or null if below threshold.
|
|
export function bestMatch(emb, voiceprints, threshold = THRESHOLD) {
|
|
let best = null, bs = -1;
|
|
for (const v of voiceprints) { const s = cosine(emb, v.vec); if (s > bs) { bs = s; best = v; } }
|
|
return (best && bs >= threshold) ? { id: best.id, name: best.name, score: bs } : null;
|
|
}
|
|
|
|
export function status() { return { model: EMB, modelId: modelId(), present: fs.existsSync(EMB), running: !!proc, threshold: THRESHOLD }; }
|