@vortex-os/computer-use 0.7.0 → 0.7.2

package/scripts/speak-supertonic.mjs

@@ -1,296 +1,296 @@
-// computer-use — Supertonic TTS speaker helper (Node + onnxruntime-node; SEPARATE INSTALL, Windows-first).
-//
-// Speaks ONE already-finalized utterance and exits — the drop-in higher-quality alternative to speak.ps1
-// (System.Speech/Heami). The CALLER (Node reflex path in mcp-stdio.mjs) owns the security: provenance
-// prefix, sanitization, and the speech budget / no-overlap. This helper just renders audio, as its own
-// short-lived process, so it never blocks the resident worker. Engine selection + Heami fallback live in
-// the caller; this script assumes models are present (the caller probes first).
-//
-// Contract (mirrors speak.ps1): --text <utt> [--voice F1] [--lang ko] [--model-dir <dir>] [--to-wav <path>]
-//   [--earcon] [--speed 1.05] [--steps 8] [--max-chars 600]. One JSON line on stdout
-//   {ok, voice, chars, ms} (or {ok:false, error}, exit 1). --to-wav renders to a file (silent, for tests).
-//
-// Inference logic is adapted from Supertone's official Node example (supertone-inc/supertonic, nodejs/helper.js,
-// MIT). Model weights (Supertone/supertonic-3) are OpenRAIL-M and are downloaded separately (fetch-supertonic.mjs),
-// never bundled. onnxruntime-node is an optionalDependency.
-
-import fs from 'node:fs';
-import os from 'node:os';
-import path from 'node:path';
-import { spawn } from 'node:child_process';
-import { fileURLToPath } from 'node:url';
-
-const __dirname = path.dirname(fileURLToPath(import.meta.url));
-
-function emit(o) { process.stdout.write(JSON.stringify(o) + '\n'); }
-
-// ── arg parse ──
-function parseArgs(argv) {
-  const a = { text: '', voice: 'F1', lang: 'ko', modelDir: '', toWav: '', earcon: false, speed: 1.05, steps: 8, maxChars: 600 };
-  for (let i = 2; i < argv.length; i++) {
-    const k = argv[i];
-    if (k === '--text') a.text = argv[++i];
-    else if (k === '--voice') a.voice = argv[++i];
-    else if (k === '--lang') a.lang = argv[++i];
-    else if (k === '--model-dir') a.modelDir = argv[++i];
-    else if (k === '--to-wav') a.toWav = argv[++i];
-    else if (k === '--earcon') a.earcon = true;
-    else if (k === '--speed') a.speed = parseFloat(argv[++i]);
-    else if (k === '--steps') a.steps = parseInt(argv[++i], 10);
-    else if (k === '--max-chars') a.maxChars = parseInt(argv[++i], 10);
-  }
-  return a;
-}
-
-// Default model dir: env override, else a per-user cache the fetch script writes to.
-function resolveModelDir(arg) {
-  if (arg) return arg;
-  if (process.env.VORTEX_CU_TTS_MODEL_DIR) return process.env.VORTEX_CU_TTS_MODEL_DIR;
-  return path.join(os.homedir(), '.vortex', 'computer-use', 'supertonic-3');
-}
-
-const AVAILABLE_LANGS = ['en', 'ko', 'ja', 'ar', 'bg', 'cs', 'da', 'de', 'el', 'es', 'et', 'fi', 'fr', 'hi', 'hr', 'hu', 'id', 'it', 'lt', 'lv', 'nl', 'pl', 'pt', 'ro', 'ru', 'sk', 'sl', 'sv', 'tr', 'uk', 'vi', 'na'];
-
-// ── text preprocessing (port of UnicodeProcessor) ──
-class UnicodeProcessor {
-  constructor(indexerPath) { this.indexer = JSON.parse(fs.readFileSync(indexerPath, 'utf8')); }
-  _pre(text, lang) {
-    text = text.normalize('NFKD');
-    const emoji = /[\u{1F600}-\u{1F64F}\u{1F300}-\u{1F5FF}\u{1F680}-\u{1F6FF}\u{1F700}-\u{1F77F}\u{1F780}-\u{1F7FF}\u{1F800}-\u{1F8FF}\u{1F900}-\u{1F9FF}\u{1FA00}-\u{1FA6F}\u{1FA70}-\u{1FAFF}\u{2600}-\u{26FF}\u{2700}-\u{27BF}\u{1F1E6}-\u{1F1FF}]+/gu;
-    text = text.replace(emoji, '');
-    const rep = { '–': '-', '‑': '-', '—': '-', '_': ' ', '“': '"', '”': '"', '‘': "'", '’': "'", '´': "'", '`': "'", '[': ' ', ']': ' ', '|': ' ', '/': ' ', '#': ' ', '→': ' ', '←': ' ' };
-    for (const [k, v] of Object.entries(rep)) text = text.replaceAll(k, v);
-    text = text.replace(/[♥☆♡©\\]/g, '');
-    const expr = { '@': ' at ', 'e.g.,': 'for example, ', 'i.e.,': 'that is, ' };
-    for (const [k, v] of Object.entries(expr)) text = text.replaceAll(k, v);
-    text = text.replace(/ ,/g, ',').replace(/ \./g, '.').replace(/ !/g, '!').replace(/ \?/g, '?').replace(/ ;/g, ';').replace(/ :/g, ':').replace(/ '/g, "'");
-    while (text.includes('""')) text = text.replace('""', '"');
-    while (text.includes("''")) text = text.replace("''", "'");
-    text = text.replace(/\s+/g, ' ').trim();
-    if (!/[.!?;:,'"')\]}…。」』】〉》›»]$/.test(text)) text += '.';
-    if (!AVAILABLE_LANGS.includes(lang)) throw new Error(`invalid lang: ${lang}`);
-    return `<${lang}>` + text + `</${lang}>`;
-  }
-  call(textList, langList) {
-    const processed = textList.map((t, i) => this._pre(t, langList[i]));
-    const lengths = processed.map((t) => t.length);
-    const maxLen = Math.max(...lengths);
-    const textIds = [];
-    for (let i = 0; i < processed.length; i++) {
-      const row = new Array(maxLen).fill(0);
-      const vals = Array.from(processed[i]).map((c) => c.charCodeAt(0));
-      for (let j = 0; j < vals.length; j++) row[j] = this.indexer[vals[j]];
-      textIds.push(row);
-    }
-    return { textIds, textMask: lengthToMask(lengths) };
-  }
-}
-
-function lengthToMask(lengths, maxLen = null) {
-  maxLen = maxLen || Math.max(...lengths);
-  return lengths.map((len) => [Array.from({ length: maxLen }, (_, j) => (j < len ? 1.0 : 0.0))]);
-}
-function getLatentMask(wavLengths, baseChunkSize, ccf) {
-  const sz = baseChunkSize * ccf;
-  return lengthToMask(wavLengths.map((len) => Math.floor((len + sz - 1) / sz)));
-}
-
-let ort; // lazy so a missing optionalDependency yields a clean JSON error, not an import crash
-function tensorF32(array, dims) { return new ort.Tensor('float32', Float32Array.from(array.flat(Infinity)), dims); }
-function tensorI64(array, dims) { return new ort.Tensor('int64', BigInt64Array.from(array.flat(Infinity).map((x) => BigInt(x))), dims); }
-
-class TextToSpeech {
-  constructor(cfgs, tp, dp, te, ve, vo) {
-    this.cfgs = cfgs; this.tp = tp; this.dp = dp; this.te = te; this.ve = ve; this.vo = vo;
-    this.sampleRate = cfgs.ae.sample_rate; this.baseChunk = cfgs.ae.base_chunk_size;
-    this.ccf = cfgs.ttl.chunk_compress_factor; this.ldim = cfgs.ttl.latent_dim;
-  }
-  _sampleNoisy(duration) {
-    const wavLenMax = Math.max(...duration) * this.sampleRate;
-    const wavLengths = duration.map((d) => Math.floor(d * this.sampleRate));
-    const chunkSize = this.baseChunk * this.ccf;
-    const latentLen = Math.floor((wavLenMax + chunkSize - 1) / chunkSize);
-    const latentDim = this.ldim * this.ccf;
-    const noisy = [];
-    for (let b = 0; b < duration.length; b++) {
-      const batch = [];
-      for (let d = 0; d < latentDim; d++) {
-        const row = [];
-        for (let t = 0; t < latentLen; t++) {
-          const u1 = Math.max(1e-10, Math.random()), u2 = Math.random();
-          row.push(Math.sqrt(-2.0 * Math.log(u1)) * Math.cos(2.0 * Math.PI * u2));
-        }
-        batch.push(row);
-      }
-      noisy.push(batch);
-    }
-    const mask = getLatentMask(wavLengths, this.baseChunk, this.ccf);
-    for (let b = 0; b < noisy.length; b++) for (let d = 0; d < noisy[b].length; d++) for (let t = 0; t < noisy[b][d].length; t++) noisy[b][d][t] *= mask[b][0][t];
-    return { noisy, mask };
-  }
-  async _infer(textList, langList, style, totalStep, speed) {
-    const bsz = textList.length;
-    const { textIds, textMask } = this.tp.call(textList, langList);
-    const idsShape = [bsz, textIds[0].length];
-    const maskTensor = tensorF32(textMask, [bsz, 1, textMask[0][0].length]);
-    const dpr = await this.dp.run({ text_ids: tensorI64(textIds, idsShape), style_dp: style.dp, text_mask: maskTensor });
-    const dur = Array.from(dpr.duration.data).map((d) => d / speed);
-    const ter = await this.te.run({ text_ids: tensorI64(textIds, idsShape), style_ttl: style.ttl, text_mask: maskTensor });
-    const textEmb = ter.text_emb;
-    let { noisy, mask } = this._sampleNoisy(dur);
-    const latShape = [bsz, noisy[0].length, noisy[0][0].length];
-    const latMaskTensor = tensorF32(mask, [bsz, 1, mask[0][0].length]);
-    const stepTensor = tensorF32(new Array(bsz).fill(totalStep), [bsz]);
-    for (let step = 0; step < totalStep; step++) {
-      const r = await this.ve.run({
-        noisy_latent: tensorF32(noisy, latShape), text_emb: textEmb, style_ttl: style.ttl,
-        text_mask: maskTensor, latent_mask: latMaskTensor, total_step: stepTensor,
-        current_step: tensorF32(new Array(bsz).fill(step), [bsz]),
-      });
-      const den = Array.from(r.denoised_latent.data);
-      let idx = 0;
-      for (let b = 0; b < noisy.length; b++) for (let d = 0; d < noisy[b].length; d++) for (let t = 0; t < noisy[b][d].length; t++) noisy[b][d][t] = den[idx++];
-    }
-    const vr = await this.vo.run({ latent: tensorF32(noisy, latShape) });
-    return { wav: Array.from(vr.wav_tts.data), duration: dur };
-  }
-  async call(text, lang, style, totalStep, speed, silence = 0.3) {
-    const maxLen = (lang === 'ko' || lang === 'ja') ? 120 : 300;
-    const chunks = chunkText(text, maxLen);
-    let wavCat = null, durCat = 0;
-    for (const chunk of chunks) {
-      const { wav, duration } = await this._infer([chunk], [lang], style, totalStep, speed);
-      if (wavCat === null) { wavCat = wav; durCat = duration[0]; }
-      else { wavCat = [...wavCat, ...new Array(Math.floor(silence * this.sampleRate)).fill(0), ...wav]; durCat += duration[0] + silence; }
-    }
-    return { wav: wavCat, duration: [durCat] };
-  }
-}
-
-function chunkText(text, maxLen) {
-  const paras = text.trim().split(/\n\s*\n+/).filter((p) => p.trim());
-  const chunks = [];
-  for (let para of paras) {
-    para = para.trim();
-    if (!para) continue;
-    const sentences = para.split(/(?<=[.!?])\s+/);
-    let cur = '';
-    for (const s of sentences) {
-      if (cur.length + s.length + 1 <= maxLen) cur += (cur ? ' ' : '') + s;
-      else { if (cur) chunks.push(cur.trim()); cur = s; }
-    }
-    if (cur) chunks.push(cur.trim());
-  }
-  return chunks.length ? chunks : [text.trim()];
-}
-
-function loadVoiceStyle(p) {
-  const vs = JSON.parse(fs.readFileSync(p, 'utf8'));
-  const td = vs.style_ttl.dims, dd = vs.style_dp.dims;
-  const ttl = new ort.Tensor('float32', Float32Array.from(vs.style_ttl.data.flat(Infinity)), [1, td[1], td[2]]);
-  const dp = new ort.Tensor('float32', Float32Array.from(vs.style_dp.data.flat(Infinity)), [1, dd[1], dd[2]]);
-  return { ttl, dp };
-}
-
-async function loadTTS(modelDir) {
-  const onnx = path.join(modelDir, 'onnx');
-  const cfgs = JSON.parse(fs.readFileSync(path.join(onnx, 'tts.json'), 'utf8'));
-  const opts = {};
-  const [dp, te, ve, vo] = await Promise.all([
-    ort.InferenceSession.create(path.join(onnx, 'duration_predictor.onnx'), opts),
-    ort.InferenceSession.create(path.join(onnx, 'text_encoder.onnx'), opts),
-    ort.InferenceSession.create(path.join(onnx, 'vector_estimator.onnx'), opts),
-    ort.InferenceSession.create(path.join(onnx, 'vocoder.onnx'), opts),
-  ]);
-  const tp = new UnicodeProcessor(path.join(onnx, 'unicode_indexer.json'));
-  return new TextToSpeech(cfgs, tp, dp, te, ve, vo);
-}
-
-// ── WAV (16-bit PCM mono) ──
-function floatToWav(samples, sampleRate) {
-  const dataSize = samples.length * 2;
-  const buf = Buffer.alloc(44 + dataSize);
-  buf.write('RIFF', 0); buf.writeUInt32LE(36 + dataSize, 4); buf.write('WAVE', 8);
-  buf.write('fmt ', 12); buf.writeUInt32LE(16, 16); buf.writeUInt16LE(1, 20); buf.writeUInt16LE(1, 22);
-  buf.writeUInt32LE(sampleRate, 24); buf.writeUInt32LE(sampleRate * 2, 28); buf.writeUInt16LE(2, 32); buf.writeUInt16LE(16, 34);
-  buf.write('data', 36); buf.writeUInt32LE(dataSize, 40);
-  for (let i = 0; i < samples.length; i++) {
-    const s = Math.max(-1, Math.min(1, samples[i]));
-    buf.writeInt16LE(Math.floor(s * 32767), 44 + i * 2);
-  }
-  return buf;
-}
-
-// Voice "slightly up": peak-normalize the synthesized voice to a consistent presence (neural TTS output often
-// sits well below full scale). Only boosts (never reduces a loud take), capped so near-silence isn't blown up.
-function normalizePeak(samples, target = 0.9, maxGain = 3.0) {
-  let peak = 0;
-  for (let i = 0; i < samples.length; i++) { const a = Math.abs(samples[i]); if (a > peak) peak = a; }
-  if (peak <= 0) return samples;
-  const gain = Math.min(target / peak, maxGain);
-  if (gain <= 1.0) return samples;
-  for (let i = 0; i < samples.length; i++) samples[i] *= gain;
-  return samples;
-}
-
-// Provenance chime (two-tone 1175→1568 Hz) matching speak.ps1's earcon, as PCM samples to prepend.
-function earconSamples(sampleRate) {
-  const tone = (freq, ms) => {
-    const n = Math.floor((ms / 1000) * sampleRate);
-    return Array.from({ length: n }, (_, i) => 0.25 * Math.sin((2 * Math.PI * freq * i) / sampleRate));
-  };
-  return [...tone(1175, 90), ...tone(1568, 110), ...new Array(Math.floor(0.06 * sampleRate)).fill(0)];
-}
-
-// Play a WAV synchronously through audio-duck.ps1 (win32): it ducks OTHER apps' sessions while the voice plays
-// and restores them in a finally, excluding its own process so the voice is never ducked. No native deps.
-// If ducking is unavailable it still plays (audio-duck falls back to a plain SoundPlayer). VORTEX_CU_DUCK=off
-// disables ducking (factor 1.0); VORTEX_CU_DUCK_FACTOR tunes how much others drop (0.5 = to 50%).
-function playWav(wavPath) {
-  return new Promise((resolve) => {
-    if (process.platform !== 'win32') return resolve(false);
-    const duckScript = path.join(__dirname, 'audio-duck.ps1');
-    const factor = process.env.VORTEX_CU_DUCK === 'off' ? '1.0' : (process.env.VORTEX_CU_DUCK_FACTOR || '0.3');
-    const ps = spawn('pwsh', ['-NoProfile', '-NonInteractive', '-File', duckScript, '-WavPath', wavPath, '-Factor', factor], { stdio: 'ignore' });
-    const kill = () => { try { ps.kill(); } catch {} };
-    process.once('SIGTERM', kill); process.once('SIGINT', kill);
-    ps.on('exit', () => resolve(true));
-    ps.on('error', () => resolve(false));
-  });
-}
-
-async function main() {
-  const a = parseArgs(process.argv);
-  let text = String(a.text || '');
-  if (text.length > a.maxChars) text = text.slice(0, a.maxChars);
-  if (!text.trim()) { emit({ ok: false, error: 'empty text' }); process.exit(1); }
-  const modelDir = resolveModelDir(a.modelDir);
-  const t0 = Date.now();
-  try {
-    ort = (await import('onnxruntime-node')).default ?? (await import('onnxruntime-node'));
-  } catch { emit({ ok: false, error: 'onnxruntime-node not installed' }); process.exit(1); }
-  try {
-    const stylePath = path.join(modelDir, 'voice_styles', `${a.voice}.json`);
-    if (!fs.existsSync(stylePath)) { emit({ ok: false, error: `voice not found: ${a.voice}` }); process.exit(1); }
-    const tts = await loadTTS(modelDir);
-    const style = loadVoiceStyle(stylePath);
-    const { wav } = await tts.call(text, a.lang, style, a.steps, a.speed);
-    const voice = normalizePeak(wav);   // voice "slightly up" — consistent presence
-    let samples = a.earcon ? [...earconSamples(tts.sampleRate), ...voice] : voice;
-    const wavBuf = floatToWav(samples, tts.sampleRate);
-    if (a.toWav) {
-      fs.writeFileSync(a.toWav, wavBuf);
-    } else {
-      const tmp = path.join(os.tmpdir(), `vortex-cu-tts-${process.pid}-${Date.now()}.wav`);
-      fs.writeFileSync(tmp, wavBuf);
-      await playWav(tmp);
-      try { fs.unlinkSync(tmp); } catch {}
-    }
-    emit({ ok: true, voice: a.voice, chars: text.length, ms: Date.now() - t0 });
-  } catch (e) {
-    emit({ ok: false, error: 'tts failed' });
-    process.exit(1);
-  }
-}
-
-main();
+// computer-use — Supertonic TTS speaker helper (Node + onnxruntime-node; SEPARATE INSTALL, Windows-first).
+//
+// Speaks ONE already-finalized utterance and exits — the drop-in higher-quality alternative to speak.ps1
+// (System.Speech/Heami). The CALLER (Node reflex path in mcp-stdio.mjs) owns the security: provenance
+// prefix, sanitization, and the speech budget / no-overlap. This helper just renders audio, as its own
+// short-lived process, so it never blocks the resident worker. Engine selection + Heami fallback live in
+// the caller; this script assumes models are present (the caller probes first).
+//
+// Contract (mirrors speak.ps1): --text <utt> [--voice F1] [--lang ko] [--model-dir <dir>] [--to-wav <path>]
+//   [--earcon] [--speed 1.05] [--steps 8] [--max-chars 600]. One JSON line on stdout
+//   {ok, voice, chars, ms} (or {ok:false, error}, exit 1). --to-wav renders to a file (silent, for tests).
+//
+// Inference logic is adapted from Supertone's official Node example (supertone-inc/supertonic, nodejs/helper.js,
+// MIT). Model weights (Supertone/supertonic-3) are OpenRAIL-M and are downloaded separately (fetch-supertonic.mjs),
+// never bundled. onnxruntime-node is an optionalDependency.
+
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import { spawn } from 'node:child_process';
+import { fileURLToPath } from 'node:url';
+
+const __dirname = path.dirname(fileURLToPath(import.meta.url));
+
+function emit(o) { process.stdout.write(JSON.stringify(o) + '\n'); }
+
+// ── arg parse ──
+function parseArgs(argv) {
+  const a = { text: '', voice: 'F1', lang: 'ko', modelDir: '', toWav: '', earcon: false, speed: 1.05, steps: 8, maxChars: 600 };
+  for (let i = 2; i < argv.length; i++) {
+    const k = argv[i];
+    if (k === '--text') a.text = argv[++i];
+    else if (k === '--voice') a.voice = argv[++i];
+    else if (k === '--lang') a.lang = argv[++i];
+    else if (k === '--model-dir') a.modelDir = argv[++i];
+    else if (k === '--to-wav') a.toWav = argv[++i];
+    else if (k === '--earcon') a.earcon = true;
+    else if (k === '--speed') a.speed = parseFloat(argv[++i]);
+    else if (k === '--steps') a.steps = parseInt(argv[++i], 10);
+    else if (k === '--max-chars') a.maxChars = parseInt(argv[++i], 10);
+  }
+  return a;
+}
+
+// Default model dir: env override, else a per-user cache the fetch script writes to.
+function resolveModelDir(arg) {
+  if (arg) return arg;
+  if (process.env.VORTEX_CU_TTS_MODEL_DIR) return process.env.VORTEX_CU_TTS_MODEL_DIR;
+  return path.join(os.homedir(), '.vortex', 'computer-use', 'supertonic-3');
+}
+
+const AVAILABLE_LANGS = ['en', 'ko', 'ja', 'ar', 'bg', 'cs', 'da', 'de', 'el', 'es', 'et', 'fi', 'fr', 'hi', 'hr', 'hu', 'id', 'it', 'lt', 'lv', 'nl', 'pl', 'pt', 'ro', 'ru', 'sk', 'sl', 'sv', 'tr', 'uk', 'vi', 'na'];
+
+// ── text preprocessing (port of UnicodeProcessor) ──
+class UnicodeProcessor {
+  constructor(indexerPath) { this.indexer = JSON.parse(fs.readFileSync(indexerPath, 'utf8')); }
+  _pre(text, lang) {
+    text = text.normalize('NFKD');
+    const emoji = /[\u{1F600}-\u{1F64F}\u{1F300}-\u{1F5FF}\u{1F680}-\u{1F6FF}\u{1F700}-\u{1F77F}\u{1F780}-\u{1F7FF}\u{1F800}-\u{1F8FF}\u{1F900}-\u{1F9FF}\u{1FA00}-\u{1FA6F}\u{1FA70}-\u{1FAFF}\u{2600}-\u{26FF}\u{2700}-\u{27BF}\u{1F1E6}-\u{1F1FF}]+/gu;
+    text = text.replace(emoji, '');
+    const rep = { '–': '-', '‑': '-', '—': '-', '_': ' ', '“': '"', '”': '"', '‘': "'", '’': "'", '´': "'", '`': "'", '[': ' ', ']': ' ', '|': ' ', '/': ' ', '#': ' ', '→': ' ', '←': ' ' };
+    for (const [k, v] of Object.entries(rep)) text = text.replaceAll(k, v);
+    text = text.replace(/[♥☆♡©\\]/g, '');
+    const expr = { '@': ' at ', 'e.g.,': 'for example, ', 'i.e.,': 'that is, ' };
+    for (const [k, v] of Object.entries(expr)) text = text.replaceAll(k, v);
+    text = text.replace(/ ,/g, ',').replace(/ \./g, '.').replace(/ !/g, '!').replace(/ \?/g, '?').replace(/ ;/g, ';').replace(/ :/g, ':').replace(/ '/g, "'");
+    while (text.includes('""')) text = text.replace('""', '"');
+    while (text.includes("''")) text = text.replace("''", "'");
+    text = text.replace(/\s+/g, ' ').trim();
+    if (!/[.!?;:,'"')\]}…。」』】〉》›»]$/.test(text)) text += '.';
+    if (!AVAILABLE_LANGS.includes(lang)) throw new Error(`invalid lang: ${lang}`);
+    return `<${lang}>` + text + `</${lang}>`;
+  }
+  call(textList, langList) {
+    const processed = textList.map((t, i) => this._pre(t, langList[i]));
+    const lengths = processed.map((t) => t.length);
+    const maxLen = Math.max(...lengths);
+    const textIds = [];
+    for (let i = 0; i < processed.length; i++) {
+      const row = new Array(maxLen).fill(0);
+      const vals = Array.from(processed[i]).map((c) => c.charCodeAt(0));
+      for (let j = 0; j < vals.length; j++) row[j] = this.indexer[vals[j]];
+      textIds.push(row);
+    }
+    return { textIds, textMask: lengthToMask(lengths) };
+  }
+}
+
+function lengthToMask(lengths, maxLen = null) {
+  maxLen = maxLen || Math.max(...lengths);
+  return lengths.map((len) => [Array.from({ length: maxLen }, (_, j) => (j < len ? 1.0 : 0.0))]);
+}
+function getLatentMask(wavLengths, baseChunkSize, ccf) {
+  const sz = baseChunkSize * ccf;
+  return lengthToMask(wavLengths.map((len) => Math.floor((len + sz - 1) / sz)));
+}
+
+let ort; // lazy so a missing optionalDependency yields a clean JSON error, not an import crash
+function tensorF32(array, dims) { return new ort.Tensor('float32', Float32Array.from(array.flat(Infinity)), dims); }
+function tensorI64(array, dims) { return new ort.Tensor('int64', BigInt64Array.from(array.flat(Infinity).map((x) => BigInt(x))), dims); }
+
+class TextToSpeech {
+  constructor(cfgs, tp, dp, te, ve, vo) {
+    this.cfgs = cfgs; this.tp = tp; this.dp = dp; this.te = te; this.ve = ve; this.vo = vo;
+    this.sampleRate = cfgs.ae.sample_rate; this.baseChunk = cfgs.ae.base_chunk_size;
+    this.ccf = cfgs.ttl.chunk_compress_factor; this.ldim = cfgs.ttl.latent_dim;
+  }
+  _sampleNoisy(duration) {
+    const wavLenMax = Math.max(...duration) * this.sampleRate;
+    const wavLengths = duration.map((d) => Math.floor(d * this.sampleRate));
+    const chunkSize = this.baseChunk * this.ccf;
+    const latentLen = Math.floor((wavLenMax + chunkSize - 1) / chunkSize);
+    const latentDim = this.ldim * this.ccf;
+    const noisy = [];
+    for (let b = 0; b < duration.length; b++) {
+      const batch = [];
+      for (let d = 0; d < latentDim; d++) {
+        const row = [];
+        for (let t = 0; t < latentLen; t++) {
+          const u1 = Math.max(1e-10, Math.random()), u2 = Math.random();
+          row.push(Math.sqrt(-2.0 * Math.log(u1)) * Math.cos(2.0 * Math.PI * u2));
+        }
+        batch.push(row);
+      }
+      noisy.push(batch);
+    }
+    const mask = getLatentMask(wavLengths, this.baseChunk, this.ccf);
+    for (let b = 0; b < noisy.length; b++) for (let d = 0; d < noisy[b].length; d++) for (let t = 0; t < noisy[b][d].length; t++) noisy[b][d][t] *= mask[b][0][t];
+    return { noisy, mask };
+  }
+  async _infer(textList, langList, style, totalStep, speed) {
+    const bsz = textList.length;
+    const { textIds, textMask } = this.tp.call(textList, langList);
+    const idsShape = [bsz, textIds[0].length];
+    const maskTensor = tensorF32(textMask, [bsz, 1, textMask[0][0].length]);
+    const dpr = await this.dp.run({ text_ids: tensorI64(textIds, idsShape), style_dp: style.dp, text_mask: maskTensor });
+    const dur = Array.from(dpr.duration.data).map((d) => d / speed);
+    const ter = await this.te.run({ text_ids: tensorI64(textIds, idsShape), style_ttl: style.ttl, text_mask: maskTensor });
+    const textEmb = ter.text_emb;
+    let { noisy, mask } = this._sampleNoisy(dur);
+    const latShape = [bsz, noisy[0].length, noisy[0][0].length];
+    const latMaskTensor = tensorF32(mask, [bsz, 1, mask[0][0].length]);
+    const stepTensor = tensorF32(new Array(bsz).fill(totalStep), [bsz]);
+    for (let step = 0; step < totalStep; step++) {
+      const r = await this.ve.run({
+        noisy_latent: tensorF32(noisy, latShape), text_emb: textEmb, style_ttl: style.ttl,
+        text_mask: maskTensor, latent_mask: latMaskTensor, total_step: stepTensor,
+        current_step: tensorF32(new Array(bsz).fill(step), [bsz]),
+      });
+      const den = Array.from(r.denoised_latent.data);
+      let idx = 0;
+      for (let b = 0; b < noisy.length; b++) for (let d = 0; d < noisy[b].length; d++) for (let t = 0; t < noisy[b][d].length; t++) noisy[b][d][t] = den[idx++];
+    }
+    const vr = await this.vo.run({ latent: tensorF32(noisy, latShape) });
+    return { wav: Array.from(vr.wav_tts.data), duration: dur };
+  }
+  async call(text, lang, style, totalStep, speed, silence = 0.3) {
+    const maxLen = (lang === 'ko' || lang === 'ja') ? 120 : 300;
+    const chunks = chunkText(text, maxLen);
+    let wavCat = null, durCat = 0;
+    for (const chunk of chunks) {
+      const { wav, duration } = await this._infer([chunk], [lang], style, totalStep, speed);
+      if (wavCat === null) { wavCat = wav; durCat = duration[0]; }
+      else { wavCat = [...wavCat, ...new Array(Math.floor(silence * this.sampleRate)).fill(0), ...wav]; durCat += duration[0] + silence; }
+    }
+    return { wav: wavCat, duration: [durCat] };
+  }
+}
+
+function chunkText(text, maxLen) {
+  const paras = text.trim().split(/\n\s*\n+/).filter((p) => p.trim());
+  const chunks = [];
+  for (let para of paras) {
+    para = para.trim();
+    if (!para) continue;
+    const sentences = para.split(/(?<=[.!?])\s+/);
+    let cur = '';
+    for (const s of sentences) {
+      if (cur.length + s.length + 1 <= maxLen) cur += (cur ? ' ' : '') + s;
+      else { if (cur) chunks.push(cur.trim()); cur = s; }
+    }
+    if (cur) chunks.push(cur.trim());
+  }
+  return chunks.length ? chunks : [text.trim()];
+}
+
+function loadVoiceStyle(p) {
+  const vs = JSON.parse(fs.readFileSync(p, 'utf8'));
+  const td = vs.style_ttl.dims, dd = vs.style_dp.dims;
+  const ttl = new ort.Tensor('float32', Float32Array.from(vs.style_ttl.data.flat(Infinity)), [1, td[1], td[2]]);
+  const dp = new ort.Tensor('float32', Float32Array.from(vs.style_dp.data.flat(Infinity)), [1, dd[1], dd[2]]);
+  return { ttl, dp };
+}
+
+async function loadTTS(modelDir) {
+  const onnx = path.join(modelDir, 'onnx');
+  const cfgs = JSON.parse(fs.readFileSync(path.join(onnx, 'tts.json'), 'utf8'));
+  const opts = {};
+  const [dp, te, ve, vo] = await Promise.all([
+    ort.InferenceSession.create(path.join(onnx, 'duration_predictor.onnx'), opts),
+    ort.InferenceSession.create(path.join(onnx, 'text_encoder.onnx'), opts),
+    ort.InferenceSession.create(path.join(onnx, 'vector_estimator.onnx'), opts),
+    ort.InferenceSession.create(path.join(onnx, 'vocoder.onnx'), opts),
+  ]);
+  const tp = new UnicodeProcessor(path.join(onnx, 'unicode_indexer.json'));
+  return new TextToSpeech(cfgs, tp, dp, te, ve, vo);
+}
+
+// ── WAV (16-bit PCM mono) ──
+function floatToWav(samples, sampleRate) {
+  const dataSize = samples.length * 2;
+  const buf = Buffer.alloc(44 + dataSize);
+  buf.write('RIFF', 0); buf.writeUInt32LE(36 + dataSize, 4); buf.write('WAVE', 8);
+  buf.write('fmt ', 12); buf.writeUInt32LE(16, 16); buf.writeUInt16LE(1, 20); buf.writeUInt16LE(1, 22);
+  buf.writeUInt32LE(sampleRate, 24); buf.writeUInt32LE(sampleRate * 2, 28); buf.writeUInt16LE(2, 32); buf.writeUInt16LE(16, 34);
+  buf.write('data', 36); buf.writeUInt32LE(dataSize, 40);
+  for (let i = 0; i < samples.length; i++) {
+    const s = Math.max(-1, Math.min(1, samples[i]));
+    buf.writeInt16LE(Math.floor(s * 32767), 44 + i * 2);
+  }
+  return buf;
+}
+
+// Voice "slightly up": peak-normalize the synthesized voice to a consistent presence (neural TTS output often
+// sits well below full scale). Only boosts (never reduces a loud take), capped so near-silence isn't blown up.
+function normalizePeak(samples, target = 0.9, maxGain = 3.0) {
+  let peak = 0;
+  for (let i = 0; i < samples.length; i++) { const a = Math.abs(samples[i]); if (a > peak) peak = a; }
+  if (peak <= 0) return samples;
+  const gain = Math.min(target / peak, maxGain);
+  if (gain <= 1.0) return samples;
+  for (let i = 0; i < samples.length; i++) samples[i] *= gain;
+  return samples;
+}
+
+// Provenance chime (two-tone 1175→1568 Hz) matching speak.ps1's earcon, as PCM samples to prepend.
+function earconSamples(sampleRate) {
+  const tone = (freq, ms) => {
+    const n = Math.floor((ms / 1000) * sampleRate);
+    return Array.from({ length: n }, (_, i) => 0.25 * Math.sin((2 * Math.PI * freq * i) / sampleRate));
+  };
+  return [...tone(1175, 90), ...tone(1568, 110), ...new Array(Math.floor(0.06 * sampleRate)).fill(0)];
+}
+
+// Play a WAV synchronously through audio-duck.ps1 (win32): it ducks OTHER apps' sessions while the voice plays
+// and restores them in a finally, excluding its own process so the voice is never ducked. No native deps.
+// If ducking is unavailable it still plays (audio-duck falls back to a plain SoundPlayer). VORTEX_CU_DUCK=off
+// disables ducking (factor 1.0); VORTEX_CU_DUCK_FACTOR tunes how much others drop (0.5 = to 50%).
+function playWav(wavPath) {
+  return new Promise((resolve) => {
+    if (process.platform !== 'win32') return resolve(false);
+    const duckScript = path.join(__dirname, 'audio-duck.ps1');
+    const factor = process.env.VORTEX_CU_DUCK === 'off' ? '1.0' : (process.env.VORTEX_CU_DUCK_FACTOR || '0.3');
+    const ps = spawn('pwsh', ['-NoProfile', '-NonInteractive', '-File', duckScript, '-WavPath', wavPath, '-Factor', factor], { stdio: 'ignore' });
+    const kill = () => { try { ps.kill(); } catch {} };
+    process.once('SIGTERM', kill); process.once('SIGINT', kill);
+    ps.on('exit', () => resolve(true));
+    ps.on('error', () => resolve(false));
+  });
+}
+
+async function main() {
+  const a = parseArgs(process.argv);
+  let text = String(a.text || '');
+  if (text.length > a.maxChars) text = text.slice(0, a.maxChars);
+  if (!text.trim()) { emit({ ok: false, error: 'empty text' }); process.exit(1); }
+  const modelDir = resolveModelDir(a.modelDir);
+  const t0 = Date.now();
+  try {
+    ort = (await import('onnxruntime-node')).default ?? (await import('onnxruntime-node'));
+  } catch { emit({ ok: false, error: 'onnxruntime-node not installed' }); process.exit(1); }
+  try {
+    const stylePath = path.join(modelDir, 'voice_styles', `${a.voice}.json`);
+    if (!fs.existsSync(stylePath)) { emit({ ok: false, error: `voice not found: ${a.voice}` }); process.exit(1); }
+    const tts = await loadTTS(modelDir);
+    const style = loadVoiceStyle(stylePath);
+    const { wav } = await tts.call(text, a.lang, style, a.steps, a.speed);
+    const voice = normalizePeak(wav);   // voice "slightly up" — consistent presence
+    let samples = a.earcon ? [...earconSamples(tts.sampleRate), ...voice] : voice;
+    const wavBuf = floatToWav(samples, tts.sampleRate);
+    if (a.toWav) {
+      fs.writeFileSync(a.toWav, wavBuf);
+    } else {
+      const tmp = path.join(os.tmpdir(), `vortex-cu-tts-${process.pid}-${Date.now()}.wav`);
+      fs.writeFileSync(tmp, wavBuf);
+      await playWav(tmp);
+      try { fs.unlinkSync(tmp); } catch {}
+    }
+    emit({ ok: true, voice: a.voice, chars: text.length, ms: Date.now() - t0 });
+  } catch (e) {
+    emit({ ok: false, error: 'tts failed' });
+    process.exit(1);
+  }
+}
+
+main();