@onjmin/koe 1.0.6 → 1.0.7

package/dist/koe-oto.js

@@ -0,0 +1,1399 @@
+#!/usr/bin/env node
+
+// src/oto/cli.ts
+import { readdir, readFile, stat, writeFile } from "fs/promises";
+import { basename, join, relative, resolve } from "path";
+
+// src/converter/wav.ts
+function parseWav(buf) {
+  const view = new DataView(buf);
+  const riff = readFourCC(view, 0);
+  if (riff !== "RIFF") throw new Error(`Not a RIFF file (got "${riff}")`);
+  let sampleRate = 0;
+  let channels = 0;
+  let bitsPerSample = 0;
+  let audioFormat = 1;
+  let dataOffset = 0;
+  let dataLength = 0;
+  let pos = 12;
+  while (pos < view.byteLength - 8) {
+    const id = readFourCC(view, pos);
+    const size = view.getUint32(pos + 4, true);
+    pos += 8;
+    if (id === "fmt ") {
+      audioFormat = view.getUint16(pos, true);
+      channels = view.getUint16(pos + 2, true);
+      sampleRate = view.getUint32(pos + 4, true);
+      bitsPerSample = view.getUint16(pos + 14, true);
+      if (audioFormat === 65534 && size >= 40) {
+        audioFormat = view.getUint16(pos + 24, true);
+      }
+    } else if (id === "data") {
+      dataOffset = pos;
+      dataLength = Math.min(size, view.byteLength - pos);
+      break;
+    }
+    pos += size + (size & 1);
+  }
+  if (!dataOffset) throw new Error("WAV has no data chunk");
+  if (!channels || !sampleRate) throw new Error("WAV fmt chunk missing");
+  const supported = audioFormat === 3 && bitsPerSample === 32 || audioFormat === 1 && (bitsPerSample === 8 || bitsPerSample === 16 || bitsPerSample === 24);
+  if (!supported) {
+    throw new Error(
+      `Unsupported WAV format ${audioFormat} / ${bitsPerSample}-bit (need PCM 8/16/24-bit or IEEE float 32-bit)`
+    );
+  }
+  const bytesPerSample = bitsPerSample >> 3;
+  const totalSamples = Math.floor(dataLength / bytesPerSample);
+  const samples = new Float32Array(totalSamples);
+  for (let i = 0; i < totalSamples; i++) {
+    const p = dataOffset + i * bytesPerSample;
+    if (audioFormat === 3) {
+      samples[i] = view.getFloat32(p, true);
+    } else if (bitsPerSample === 8) {
+      samples[i] = (view.getUint8(p) - 128) / 128;
+    } else if (bitsPerSample === 16) {
+      samples[i] = view.getInt16(p, true) / 32768;
+    } else if (bitsPerSample === 24) {
+      const lo = view.getUint8(p) | view.getUint8(p + 1) << 8;
+      let hi = view.getUint8(p + 2);
+      if (hi & 128) hi = hi | 4294967040;
+      samples[i] = (hi << 16 | lo) / 8388608;
+    }
+  }
+  return { sampleRate, channels, samples };
+}
+function toMono(wav) {
+  if (wav.channels === 1) return wav;
+  const len = wav.samples.length / wav.channels;
+  const out = new Float32Array(len);
+  for (let i = 0; i < len; i++) {
+    let sum = 0;
+    for (let c = 0; c < wav.channels; c++)
+      sum += wav.samples[i * wav.channels + c];
+    out[i] = sum / wav.channels;
+  }
+  return { sampleRate: wav.sampleRate, channels: 1, samples: out };
+}
+function resample(wav, targetRate) {
+  if (wav.sampleRate === targetRate) return wav;
+  const ratio = wav.sampleRate / targetRate;
+  const outLen = Math.floor(wav.samples.length / ratio);
+  const out = new Float32Array(outLen);
+  const src = wav.samples;
+  for (let i = 0; i < outLen; i++) {
+    const x = i * ratio;
+    const xi = Math.floor(x);
+    const frac = x - xi;
+    out[i] = (src[xi] ?? 0) + ((src[xi + 1] ?? 0) - (src[xi] ?? 0)) * frac;
+  }
+  return { sampleRate: targetRate, channels: 1, samples: out };
+}
+function readFourCC(view, pos) {
+  return String.fromCharCode(
+    view.getUint8(pos),
+    view.getUint8(pos + 1),
+    view.getUint8(pos + 2),
+    view.getUint8(pos + 3)
+  );
+}
+
+// src/oto/frames.ts
+var RATE = 16e3;
+var HOP_MS = 2;
+var HOP = RATE * HOP_MS / 1e3;
+var FFT_SIZE = 512;
+var BINS = FFT_SIZE / 2;
+var PITCH_DECIM = 2;
+var PITCH_RATE = RATE / PITCH_DECIM;
+var PITCH_WIN = 512;
+var MIN_F0 = 70;
+var MAX_F0 = 600;
+var WINDOW_CENTRE_MS = FFT_SIZE / 2 / RATE * 1e3;
+function msToFrame(ms) {
+  return Math.round(ms / HOP_MS);
+}
+function framesToMs(frames) {
+  return frames * HOP_MS;
+}
+function frameTimeMs(frame) {
+  return frame * HOP_MS + WINDOW_CENTRE_MS;
+}
+function fft(re, im) {
+  const n = re.length;
+  for (let i = 1, j = 0; i < n; i++) {
+    let bit = n >> 1;
+    for (; j & bit; bit >>= 1) j ^= bit;
+    j ^= bit;
+    if (i < j) {
+      [re[i], re[j]] = [re[j], re[i]];
+      [im[i], im[j]] = [im[j], im[i]];
+    }
+  }
+  for (let len = 2; len <= n; len <<= 1) {
+    const ang = -2 * Math.PI / len;
+    const wRe = Math.cos(ang);
+    const wIm = Math.sin(ang);
+    for (let i = 0; i < n; i += len) {
+      let curRe = 1;
+      let curIm = 0;
+      for (let k = 0; k < len >> 1; k++) {
+        const aRe = re[i + k];
+        const aIm = im[i + k];
+        const bRe = re[i + k + (len >> 1)] * curRe - im[i + k + (len >> 1)] * curIm;
+        const bIm = re[i + k + (len >> 1)] * curIm + im[i + k + (len >> 1)] * curRe;
+        re[i + k] = aRe + bRe;
+        im[i + k] = aIm + bIm;
+        re[i + k + (len >> 1)] = aRe - bRe;
+        im[i + k + (len >> 1)] = aIm - bIm;
+        const nextRe = curRe * wRe - curIm * wIm;
+        curIm = curRe * wIm + curIm * wRe;
+        curRe = nextRe;
+      }
+    }
+  }
+}
+function quietestWindow(db) {
+  const averaged = smooth(db, Math.round(50 / HOP_MS));
+  let min = Infinity;
+  for (let i = 0; i < averaged.length; i++) {
+    if (averaged[i] < min) min = averaged[i];
+  }
+  return Number.isFinite(min) ? min : -120;
+}
+function percentile(values, p) {
+  const sorted = Float32Array.from(values).sort();
+  const i = Math.min(
+    sorted.length - 1,
+    Math.max(0, Math.round(p * (sorted.length - 1)))
+  );
+  return sorted[i];
+}
+function ncc(buf, at, win, lag) {
+  let r = 0;
+  let e0 = 0;
+  let e1 = 0;
+  for (let i = 0; i < win; i++) {
+    const a = buf[at + i];
+    const b = buf[at + i + lag];
+    r += a * b;
+    e0 += a * a;
+    e1 += b * b;
+  }
+  const denom = Math.sqrt(e0 * e1);
+  return denom > 1e-12 ? r / denom : 0;
+}
+function analyze(wav) {
+  const mono = resample(toMono(wav), RATE);
+  const x = mono.samples;
+  const n = Math.max(1, Math.floor((x.length - FFT_SIZE) / HOP) + 1);
+  const rmsDb = new Float32Array(n);
+  const highRatio = new Float32Array(n);
+  const highDb = new Float32Array(n);
+  const flux = new Float32Array(n);
+  const voiced = new Float32Array(n);
+  const window = new Float32Array(FFT_SIZE);
+  for (let i = 0; i < FFT_SIZE; i++) {
+    window[i] = 0.5 - 0.5 * Math.cos(2 * Math.PI * i / FFT_SIZE);
+  }
+  const re = new Float32Array(FFT_SIZE);
+  const im = new Float32Array(FFT_SIZE);
+  const mag = new Float32Array(BINS);
+  const prevMag = new Float32Array(BINS);
+  const highBin = Math.floor(4e3 / (RATE / 2) * BINS);
+  for (let t = 0; t < n; t++) {
+    const off = t * HOP;
+    let energy = 0;
+    for (let i = 0; i < FFT_SIZE; i++) {
+      const s = x[off + i] ?? 0;
+      energy += s * s;
+      re[i] = s * window[i];
+      im[i] = 0;
+    }
+    rmsDb[t] = 10 * Math.log10(energy / FFT_SIZE + 1e-12);
+    fft(re, im);
+    let lowSum = 0;
+    let highSum = 0;
+    let fluxSum = 0;
+    for (let b = 1; b < BINS; b++) {
+      const m = Math.sqrt(re[b] * re[b] + im[b] * im[b]);
+      mag[b] = m;
+      if (b < highBin) lowSum += m;
+      else highSum += m;
+      const d = Math.log10(m + 1e-6) - Math.log10(prevMag[b] + 1e-6);
+      if (d > 0) fluxSum += d;
+    }
+    prevMag.set(mag);
+    highRatio[t] = highSum / (lowSum + highSum + 1e-12);
+    highDb[t] = 20 * Math.log10(highSum / BINS + 1e-9);
+    flux[t] = t === 0 ? 0 : fluxSum;
+  }
+  const smoothDb = smooth(rmsDb, Math.round(30 / HOP_MS));
+  const smoothHighDb = smooth(highDb, Math.round(20 / HOP_MS));
+  const highFloorDb = quietestWindow(smoothHighDb);
+  const floorDb = percentile(smoothDb, 0.1);
+  const quietDb = quietestWindow(smoothDb);
+  let peakDb = -120;
+  for (let t = 0; t < n; t++) if (smoothDb[t] > peakDb) peakDb = smoothDb[t];
+  const pitchBuf = new Float32Array(Math.floor(x.length / PITCH_DECIM));
+  for (let i = 0; i < pitchBuf.length; i++) {
+    pitchBuf[i] = (x[i * PITCH_DECIM] + x[i * PITCH_DECIM + 1]) * 0.5;
+  }
+  const minLag = Math.floor(PITCH_RATE / MAX_F0);
+  const maxLag = Math.floor(PITCH_RATE / MIN_F0);
+  const loud = [];
+  for (let t = 0; t < n; t++) {
+    if (smoothDb[t] > peakDb - 10) loud.push(t);
+  }
+  const lags = [];
+  const step = Math.max(1, Math.floor(loud.length / 24));
+  for (let k = 0; k < loud.length; k += step) {
+    const at = Math.floor(loud[k] * HOP / PITCH_DECIM);
+    if (at + PITCH_WIN + maxLag >= pitchBuf.length) continue;
+    let best = 0;
+    let bestLag = 0;
+    for (let lag = minLag; lag <= maxLag; lag++) {
+      const v = ncc(pitchBuf, at, PITCH_WIN, lag);
+      if (v > best) {
+        best = v;
+        bestLag = lag;
+      }
+    }
+    if (best > 0.5 && bestLag > 0) lags.push(bestLag);
+  }
+  lags.sort((a, b) => a - b);
+  const centreLag = lags.length ? lags[lags.length >> 1] : 0;
+  const f0 = centreLag ? PITCH_RATE / centreLag : 0;
+  if (centreLag) {
+    const lo = Math.max(minLag, Math.floor(centreLag / 1.7));
+    const hi = Math.min(maxLag, Math.ceil(centreLag * 1.7));
+    for (let t = 0; t < n; t++) {
+      if (smoothDb[t] < floorDb + 6) continue;
+      const at = Math.floor(t * HOP / PITCH_DECIM);
+      if (at + PITCH_WIN + hi >= pitchBuf.length) continue;
+      let best = 0;
+      for (let lag = lo; lag <= hi; lag++) {
+        const v = ncc(pitchBuf, at, PITCH_WIN, lag);
+        if (v > best) best = v;
+      }
+      voiced[t] = best;
+    }
+  }
+  const medianFlux = percentile(flux, 0.5) || 1;
+  for (let t = 0; t < n; t++) flux[t] /= medianFlux;
+  return {
+    n,
+    durationMs: x.length / RATE * 1e3,
+    rmsDb,
+    smoothDb,
+    voiced,
+    highRatio,
+    highDb: smoothHighDb,
+    highFloorDb,
+    flux: smooth(flux, 3),
+    f0,
+    floorDb,
+    quietDb,
+    peakDb
+  };
+}
+function smooth(src, width) {
+  if (width <= 1) return Float32Array.from(src);
+  const prefix = new Float64Array(src.length + 1);
+  for (let i = 0; i < src.length; i++) prefix[i + 1] = prefix[i] + src[i];
+  const out = new Float32Array(src.length);
+  const half = width >> 1;
+  for (let i = 0; i < src.length; i++) {
+    const lo = Math.max(0, i - half);
+    const hi = Math.min(src.length, i + half + 1);
+    out[i] = (prefix[hi] - prefix[lo]) / (hi - lo);
+  }
+  return out;
+}
+
+// src/oto/estimate.ts
+var ARTICULATION = {
+  vowel: {
+    maxConsonantMs: 40,
+    leadInMs: 3,
+    overlapRatio: 0,
+    fixedOverlapMs: 20,
+    voicedConsonant: true,
+    fricationOnset: false
+  },
+  nasalN: {
+    maxConsonantMs: 40,
+    leadInMs: 3,
+    overlapRatio: 0,
+    fixedOverlapMs: 20,
+    voicedConsonant: true,
+    fricationOnset: false
+  },
+  nasal: {
+    maxConsonantMs: 110,
+    leadInMs: 5,
+    overlapRatio: 0.6,
+    voicedConsonant: true,
+    fricationOnset: false
+  },
+  liquid: {
+    maxConsonantMs: 80,
+    leadInMs: 5,
+    overlapRatio: 0.6,
+    voicedConsonant: true,
+    fricationOnset: false
+  },
+  semivowel: {
+    maxConsonantMs: 75,
+    leadInMs: 5,
+    overlapRatio: 0.6,
+    voicedConsonant: true,
+    fricationOnset: false
+  },
+  fricativeVoiceless: {
+    maxConsonantMs: 140,
+    leadInMs: 5,
+    overlapRatio: 0.3,
+    voicedConsonant: false,
+    fricationOnset: true
+  },
+  fricativeVoiced: {
+    maxConsonantMs: 130,
+    leadInMs: 5,
+    overlapRatio: 0.4,
+    voicedConsonant: true,
+    fricationOnset: true
+  },
+  affricate: {
+    maxConsonantMs: 120,
+    leadInMs: 5,
+    overlapRatio: 0.3,
+    voicedConsonant: false,
+    fricationOnset: true
+  },
+  plosiveVoiceless: {
+    maxConsonantMs: 90,
+    leadInMs: 8,
+    overlapRatio: 0,
+    // か/た/ぱ行 need a gap, not a crossfade: the wiki asks for a negative
+    // overlap so the closure that precedes the burst survives synthesis.
+    fixedOverlapMs: -10,
+    voicedConsonant: false,
+    fricationOnset: false
+  },
+  plosiveVoiced: {
+    maxConsonantMs: 90,
+    leadInMs: 8,
+    overlapRatio: 0.55,
+    voicedConsonant: true,
+    fricationOnset: false
+  }
+};
+var MIN_OVERLAP_MS = 12;
+var MAX_OVERLAP_MS = 40;
+var VOICED_THRESHOLD = 0.55;
+var VOICED_RUN_MS = 24;
+var MIN_VOICED_CONSONANT_MS = 22;
+var SEQUENCE_ONSET_BIAS = 12;
+var MAX_SEQUENCE_PRE_MS = 250;
+var DECAY_DROP_DB = 4;
+var RELEASE_GUARD_FRAMES = 10;
+function clamp(v, lo, hi) {
+  return Math.min(hi, Math.max(lo, v));
+}
+var SOUND_OVER_FLOOR_DB = 8;
+var SOUND_UNDER_PEAK_DB = 15;
+var ATTACK_SLOPE_DB = 1;
+var FRICATION_HIGH_RATIO = 0.45;
+var FRICATION_UNDER_PEAK_DB = 30;
+function soundThreshold(f) {
+  return Math.min(
+    Math.max(f.floorDb + SOUND_OVER_FLOOR_DB, f.peakDb - 34),
+    f.peakDb - SOUND_UNDER_PEAK_DB
+  );
+}
+function findSoundStart(f, from, to) {
+  const th = soundThreshold(f);
+  for (let t = Math.max(0, from); t < Math.min(f.n, to); t++) {
+    if (f.smoothDb[t] >= th) return t;
+  }
+  return Math.max(0, from);
+}
+function findSoundEnd(f, from, to) {
+  const th = soundThreshold(f);
+  for (let t = Math.min(f.n, to) - 1; t > from; t--) {
+    if (f.smoothDb[t] >= th) return t;
+  }
+  return Math.min(f.n - 1, to);
+}
+function findConsonantStart(f, from, to, frication) {
+  const lo = Math.max(0, from);
+  let start = findSoundStart(f, from, to);
+  if (frication) {
+    const level = f.peakDb - FRICATION_UNDER_PEAK_DB;
+    const absolute = f.highFloorDb + 8;
+    for (let t = lo; t < start; t++) {
+      const byRatio = f.highRatio[t] >= FRICATION_HIGH_RATIO && f.smoothDb[t] >= level;
+      if (byRatio || f.highDb[t] >= absolute) {
+        start = t;
+        break;
+      }
+    }
+  }
+  const foot = f.quietDb + 6;
+  const k = Math.max(1, msToFrame(6));
+  while (start > lo) {
+    const back = Math.max(lo, start - k);
+    const stillFalling = f.smoothDb[start] - f.smoothDb[back] >= ATTACK_SLOPE_DB;
+    if (!stillFalling && f.smoothDb[start - 1] <= foot) break;
+    start--;
+  }
+  return start;
+}
+function findVoiceOnset(f, from, to) {
+  const need = Math.ceil(VOICED_RUN_MS / HOP_MS);
+  const hi = Math.min(f.n, to);
+  for (let t = Math.max(0, from); t < hi; t++) {
+    if (f.voiced[t] < VOICED_THRESHOLD) continue;
+    let run = 0;
+    while (t + run < hi && f.voiced[t + run] >= VOICED_THRESHOLD) run++;
+    if (run >= need) return t;
+    t += run;
+  }
+  return -1;
+}
+function findVoicedRelease(f, from, to) {
+  const lo = Math.max(1, from);
+  const hi = Math.min(f.n - 1, to);
+  if (hi <= lo) return lo;
+  const k = Math.max(1, msToFrame(10));
+  let best = lo;
+  let bestScore = -Infinity;
+  for (let t = lo; t < hi; t++) {
+    const rise = f.smoothDb[Math.min(f.n - 1, t + k)] - f.smoothDb[Math.max(0, t - k)];
+    const deSibilance = (f.highRatio[Math.max(0, t - k)] - f.highRatio[Math.min(f.n - 1, t + k)]) * 30;
+    const score = rise + deSibilance + f.flux[t];
+    if (score > bestScore) {
+      bestScore = score;
+      best = t;
+    }
+  }
+  return best;
+}
+function findVowelStable(f, vowelOnset, limit) {
+  const minMs = 45;
+  const maxMs = 110;
+  const from = vowelOnset + msToFrame(minMs);
+  const to = Math.min(limit, vowelOnset + msToFrame(maxMs));
+  if (to <= from) return Math.min(limit, vowelOnset + msToFrame(minMs));
+  const calm = smooth(f.flux, Math.max(1, msToFrame(12)));
+  const need = Math.ceil(20 / HOP_MS);
+  for (let t = from; t < to - need; t++) {
+    let steady = true;
+    for (let k = 0; k < need; k++) {
+      if (calm[t + k] > 1.05) {
+        steady = false;
+        break;
+      }
+    }
+    if (steady) return t;
+  }
+  return to;
+}
+function findDecayStart(f, from, soundEnd) {
+  let peak = from;
+  for (let t = from; t <= soundEnd; t++) {
+    if (f.smoothDb[t] > f.smoothDb[peak]) peak = t;
+  }
+  const th = f.smoothDb[peak] - DECAY_DROP_DB;
+  const need = Math.ceil(24 / HOP_MS);
+  for (let t = peak; t <= soundEnd - need; t++) {
+    let falling = true;
+    for (let k = 0; k < need; k++) {
+      if (f.smoothDb[t + k] >= th) {
+        falling = false;
+        break;
+      }
+    }
+    if (falling) return Math.max(from, t - RELEASE_GUARD_FRAMES);
+  }
+  return Math.max(from, soundEnd - RELEASE_GUARD_FRAMES);
+}
+function hasPrevoicing(f, consStart, vowelOnset) {
+  if (vowelOnset - consStart < msToFrame(25)) return false;
+  let voicedFrames = 0;
+  for (let t = consStart; t < vowelOnset; t++) {
+    if (f.voiced[t] >= VOICED_THRESHOLD) voicedFrames++;
+  }
+  return voicedFrames >= (vowelOnset - consStart) * 0.6;
+}
+function locateMora(f, cls, from, to) {
+  const art = ARTICULATION[cls];
+  const consStart = findConsonantStart(f, from, to, art.fricationOnset);
+  const searchEnd = Math.min(
+    to,
+    consStart + msToFrame(art.maxConsonantMs + 60)
+  );
+  let vowelOnset;
+  if (!art.voicedConsonant) {
+    const v = findVoiceOnset(f, consStart, searchEnd);
+    vowelOnset = v >= 0 ? v : consStart + msToFrame(30);
+  } else if (cls === "vowel" || cls === "nasalN") {
+    const v = findVoiceOnset(f, consStart, searchEnd);
+    vowelOnset = v >= 0 ? v : consStart;
+  } else {
+    const voiceStart = findVoiceOnset(f, consStart, searchEnd);
+    const base = voiceStart >= 0 ? voiceStart : consStart;
+    vowelOnset = findVoicedRelease(
+      f,
+      base + msToFrame(MIN_VOICED_CONSONANT_MS),
+      Math.min(to, base + msToFrame(art.maxConsonantMs))
+    );
+  }
+  vowelOnset = clamp(vowelOnset, consStart, Math.max(consStart, to - 1));
+  return {
+    consStart,
+    vowelOnset,
+    stable: findVowelStable(f, vowelOnset, to),
+    prevoiced: cls === "plosiveVoiced" && hasPrevoicing(f, consStart, vowelOnset)
+  };
+}
+function round(ms) {
+  return Math.round(ms * 1e3) / 1e3;
+}
+function buildEntry(wav, alias, f, cls, pos, endFrame) {
+  const art = ARTICULATION[cls];
+  let offsetMs = frameTimeMs(pos.consStart) - art.leadInMs;
+  const vowelMs = frameTimeMs(pos.vowelOnset);
+  offsetMs = Math.max(offsetMs, vowelMs - art.maxConsonantMs);
+  offsetMs = clamp(offsetMs, 0, Math.max(0, vowelMs - 1));
+  const pre = Math.max(1, vowelMs - offsetMs);
+  let overlap;
+  if (art.fixedOverlapMs !== void 0) {
+    overlap = art.fixedOverlapMs;
+  } else {
+    overlap = clamp(
+      art.overlapRatio * pre,
+      Math.min(MIN_OVERLAP_MS, pre),
+      Math.min(MAX_OVERLAP_MS, pre)
+    );
+  }
+  const consonant = Math.max(
+    pre + 10,
+    Math.abs(overlap) + 10,
+    frameTimeMs(pos.stable) - offsetMs
+  );
+  const endMs = Math.max(frameTimeMs(endFrame), offsetMs + consonant + 30);
+  const cutoff = -(Math.min(endMs, f.durationMs) - offsetMs);
+  return {
+    wav,
+    alias,
+    offset: round(offsetMs),
+    consonant: round(consonant),
+    cutoff: round(cutoff),
+    pre: round(pre),
+    overlap: round(overlap)
+  };
+}
+function estimateSolo(wav, f, syl, aliases) {
+  const soundStart = findSoundStart(f, 0, f.n);
+  const soundEnd = findSoundEnd(f, soundStart, f.n);
+  const pos = locateMora(f, syl.cls, 0, soundEnd);
+  const end = findDecayStart(f, pos.stable, soundEnd);
+  return aliases.map((alias) => buildEntry(wav, alias, f, syl.cls, pos, end));
+}
+function estimateVowelJoin(wav, f, syl, alias) {
+  const soundStart = findSoundStart(f, 0, f.n);
+  const soundEnd = findSoundEnd(f, soundStart, f.n);
+  const pos = locateMora(f, syl.cls, 0, soundEnd);
+  const end = findDecayStart(f, pos.stable, soundEnd);
+  const pre = 50;
+  const overlap = 100;
+  const consonant = 110;
+  const offsetMs = frameTimeMs(pos.stable);
+  const endMs = frameTimeMs(end);
+  if (offsetMs < 0 || endMs - offsetMs < consonant + 40) return null;
+  return {
+    wav,
+    alias,
+    offset: round(offsetMs),
+    consonant,
+    cutoff: round(-(endMs - offsetMs)),
+    pre,
+    overlap
+  };
+}
+var GRID_SNAP_WEIGHT = 0.6;
+function fitLine(values) {
+  const n = values.length;
+  if (n < 2) return { intercept: values[0] ?? 0, slope: 0 };
+  let sx = 0;
+  let sy = 0;
+  let sxx = 0;
+  let sxy = 0;
+  for (let i = 0; i < n; i++) {
+    sx += i;
+    sy += values[i];
+    sxx += i * i;
+    sxy += i * values[i];
+  }
+  const denom = n * sxx - sx * sx;
+  const slope = denom !== 0 ? (n * sxy - sx * sy) / denom : 0;
+  return { intercept: (sy - slope * sx) / n, slope };
+}
+function onsetStrength(f) {
+  const k = Math.max(1, msToFrame(12));
+  const out = new Float32Array(f.n);
+  for (let t = 0; t < f.n; t++) {
+    const rise = f.smoothDb[Math.min(f.n - 1, t + k)] - f.smoothDb[Math.max(0, t - k)];
+    out[t] = f.flux[t] + Math.max(0, rise) / 3;
+  }
+  return smooth(out, Math.max(1, msToFrame(8)));
+}
+function detectGrid(f, count) {
+  if (count < 1) return null;
+  const strength = onsetStrength(f);
+  const uttStart = findSoundStart(f, 0, f.n);
+  const uttEnd = findSoundEnd(f, uttStart, f.n);
+  const span = uttEnd - uttStart;
+  if (span < msToFrame(80)) return null;
+  if (count === 1) {
+    return { start: uttStart, interval: span, onsets: [uttStart] };
+  }
+  const candidates = intervalCandidates(
+    strength,
+    uttStart,
+    uttEnd,
+    span / count
+  );
+  const best = { score: -Infinity, start: uttStart, interval: span / count };
+  const search = (requireFit) => {
+    for (const candidate of candidates) {
+      for (let iv = candidate * 0.85; iv <= candidate * 1.18; iv += candidate * 0.02) {
+        const near2 = windowArgMax(strength, Math.round(iv * 0.22));
+        const tol = iv * 0.22;
+        const hardEnd = uttEnd + iv * 0.4;
+        for (let phase = uttStart - iv * 0.25; phase <= uttStart + iv * 0.5; phase += 1) {
+          if (requireFit && phase + (count - 1) * iv > hardEnd) break;
+          let total = 0;
+          for (let i = 0; i < count; i++) {
+            const centre = phase + i * iv;
+            const at = near2[clamp(Math.round(centre), 0, f.n - 1)];
+            total += strength[at] - Math.abs(at - centre) / tol;
+          }
+          if (total > best.score) {
+            best.score = total;
+            best.start = phase;
+            best.interval = iv;
+          }
+        }
+      }
+    }
+  };
+  search(true);
+  if (!Number.isFinite(best.score)) search(false);
+  if (!Number.isFinite(best.score)) return null;
+  const near = windowArgMax(strength, Math.round(best.interval * 0.25));
+  const snapped = [];
+  for (let i = 0; i < count; i++) {
+    snapped.push(
+      near[clamp(Math.round(best.start + i * best.interval), 0, f.n - 1)]
+    );
+  }
+  const { intercept, slope } = fitLine(snapped);
+  const onsets = [];
+  for (let i = 0; i < count; i++) {
+    const fitted = intercept + slope * i;
+    const blended = Math.round(
+      fitted + (snapped[i] - fitted) * GRID_SNAP_WEIGHT
+    );
+    onsets.push(i > 0 ? Math.max(blended, onsets[i - 1] + 4) : blended);
+  }
+  return {
+    start: onsets[0],
+    interval: slope > 0 ? slope : best.interval,
+    onsets
+  };
+}
+function windowArgMax(src, half) {
+  const n = src.length;
+  const out = new Int32Array(n);
+  const deque = new Int32Array(n);
+  let head = 0;
+  let tail = 0;
+  let next = 0;
+  for (let i = 0; i < n; i++) {
+    const limit = Math.min(n - 1, i + half);
+    while (next <= limit) {
+      while (tail > head && src[deque[tail - 1]] <= src[next]) tail--;
+      deque[tail++] = next++;
+    }
+    while (deque[head] < i - half) head++;
+    out[i] = deque[head];
+  }
+  return out;
+}
+function intervalCandidates(strength, from, to, fromSpan) {
+  const lo = msToFrame(180);
+  const hi = Math.min(msToFrame(1300), Math.floor((to - from) / 2));
+  const out = [];
+  if (hi > lo) {
+    let mean = 0;
+    for (let t = from; t < to; t++) mean += strength[t];
+    mean /= Math.max(1, to - from);
+    const r = new Float64Array(hi + 1);
+    let peak = 0;
+    for (let lag = lo; lag <= hi; lag++) {
+      let sum = 0;
+      let n = 0;
+      for (let t = from; t + lag < to; t++, n++) {
+        sum += (strength[t] - mean) * (strength[t + lag] - mean);
+      }
+      r[lag] = n > 0 ? sum / n : 0;
+      if (r[lag] > peak) peak = r[lag];
+    }
+    if (peak > 0) {
+      for (let lag = lo + 1; lag < hi; lag++) {
+        if (r[lag] < peak * 0.7) continue;
+        if (r[lag] < r[lag - 1] || r[lag] < r[lag + 1]) continue;
+        out.push(lag);
+        if (out.length >= 2) break;
+      }
+    }
+  }
+  out.push(fromSpan);
+  const withSubmultiples = [];
+  const seen = /* @__PURE__ */ new Set();
+  const shortest = msToFrame(150);
+  for (const c of out) {
+    for (const divisor of [1, 2, 3]) {
+      const key = Math.round(c / divisor);
+      if (key < shortest || seen.has(key)) continue;
+      seen.add(key);
+      withSubmultiples.push(key);
+    }
+  }
+  if (withSubmultiples.length === 0) withSubmultiples.push(shortest);
+  return withSubmultiples;
+}
+function sequenceVowelOnset(f, cls, onset, next) {
+  if (ARTICULATION[cls].voicedConsonant) return onset + SEQUENCE_ONSET_BIAS;
+  const lo = Math.max(0, onset - msToFrame(100));
+  const hi = Math.min(next, onset + msToFrame(150));
+  let quietest = onset;
+  for (let t = lo; t < hi; t++) {
+    if (f.voiced[t] < f.voiced[quietest]) quietest = t;
+  }
+  if (f.voiced[quietest] >= VOICED_THRESHOLD) return onset;
+  const resumed = findVoiceOnset(
+    f,
+    quietest,
+    Math.min(next, hi + msToFrame(80))
+  );
+  return resumed >= 0 ? resumed : onset;
+}
+function estimateSequence(wav, f, syllables, opts = {}) {
+  const grid = detectGrid(f, syllables.length);
+  if (!grid) return [];
+  const intervalMs = framesToMs(grid.interval);
+  const pre = clamp(intervalMs / 2, 60, MAX_SEQUENCE_PRE_MS);
+  const overlap = round(pre / 3);
+  const consonant = round(pre * 1.5);
+  const suffix = opts.suffix ?? "";
+  const prefix = opts.prefix ?? "";
+  const entries = [];
+  for (let i = 0; i < syllables.length; i++) {
+    const syl = syllables[i];
+    const onset = grid.onsets[i];
+    const next = i + 1 < syllables.length ? grid.onsets[i + 1] : findSoundEnd(f, onset, f.n);
+    let vowelFrame;
+    if (i === 0) {
+      vowelFrame = locateMora(f, syl.cls, 0, next).vowelOnset;
+    } else {
+      vowelFrame = sequenceVowelOnset(f, syl.cls, onset, next);
+    }
+    const noteMs = frameTimeMs(vowelFrame);
+    const offsetMs = clamp(noteMs - pre, 0, Math.max(0, noteMs - 1));
+    const actualPre = noteMs - offsetMs;
+    const endMs = Math.min(f.durationMs, noteMs + intervalMs * 2 / 3);
+    const alias = i === 0 ? `- ${prefix}${syl.kana}${suffix}` : `${syllables[i - 1].vowel} ${syl.kana}${suffix}`;
+    entries.push({
+      wav,
+      alias,
+      offset: round(offsetMs),
+      consonant: Math.max(round(consonant), round(actualPre + 20)),
+      cutoff: round(-Math.max(endMs - offsetMs, consonant + 40)),
+      pre: round(actualPre),
+      overlap: i === 0 ? 0 : Math.min(overlap, round(actualPre))
+    });
+  }
+  if (opts.trailingRest && syllables.length > 0) {
+    const last = syllables[syllables.length - 1];
+    const lastOnset = grid.onsets[syllables.length - 1];
+    const soundEnd = findSoundEnd(f, lastOnset, f.n);
+    const decay = findDecayStart(f, lastOnset, soundEnd);
+    const noteMs = frameTimeMs(decay);
+    const offsetMs = clamp(noteMs - pre, 0, Math.max(0, noteMs - 1));
+    const actualPre = noteMs - offsetMs;
+    entries.push({
+      wav,
+      alias: `${last.vowel} R${suffix}`,
+      offset: round(offsetMs),
+      consonant: round(consonant),
+      cutoff: round(
+        -(Math.min(f.durationMs, frameTimeMs(soundEnd)) - offsetMs)
+      ),
+      pre: round(actualPre),
+      overlap: Math.min(overlap, round(actualPre))
+    });
+  }
+  return entries;
+}
+
+// src/oto/kana.ts
+var CLASS_OF = {
+  "": "vowel",
+  k: "plosiveVoiceless",
+  ky: "plosiveVoiceless",
+  t: "plosiveVoiceless",
+  ty: "plosiveVoiceless",
+  p: "plosiveVoiceless",
+  py: "plosiveVoiceless",
+  g: "plosiveVoiced",
+  gy: "plosiveVoiced",
+  d: "plosiveVoiced",
+  dy: "plosiveVoiced",
+  b: "plosiveVoiced",
+  by: "plosiveVoiced",
+  s: "fricativeVoiceless",
+  sh: "fricativeVoiceless",
+  h: "fricativeVoiceless",
+  hy: "fricativeVoiceless",
+  f: "fricativeVoiceless",
+  z: "fricativeVoiced",
+  j: "fricativeVoiced",
+  v: "fricativeVoiced",
+  ts: "affricate",
+  ch: "affricate",
+  n: "nasal",
+  ny: "nasal",
+  m: "nasal",
+  my: "nasal",
+  r: "liquid",
+  ry: "liquid",
+  y: "semivowel",
+  w: "semivowel"
+};
+var KANA = {
+  \u3042: " a",
+  \u3044: " i",
+  \u3046: " u",
+  \u3048: " e",
+  \u304A: " o",
+  \u3093: " n",
+  \u304B: "k a",
+  \u304D: "k i",
+  \u304F: "k u",
+  \u3051: "k e",
+  \u3053: "k o",
+  \u304C: "g a",
+  \u304E: "g i",
+  \u3050: "g u",
+  \u3052: "g e",
+  \u3054: "g o",
+  \u3055: "s a",
+  \u3057: "sh i",
+  \u3059: "s u",
+  \u305B: "s e",
+  \u305D: "s o",
+  \u3056: "z a",
+  \u3058: "j i",
+  \u305A: "z u",
+  \u305C: "z e",
+  \u305E: "z o",
+  \u305F: "t a",
+  \u3061: "ch i",
+  \u3064: "ts u",
+  \u3066: "t e",
+  \u3068: "t o",
+  \u3060: "d a",
+  \u3062: "j i",
+  \u3065: "z u",
+  \u3067: "d e",
+  \u3069: "d o",
+  \u306A: "n a",
+  \u306B: "n i",
+  \u306C: "n u",
+  \u306D: "n e",
+  \u306E: "n o",
+  \u306F: "h a",
+  \u3072: "h i",
+  \u3075: "f u",
+  \u3078: "h e",
+  \u307B: "h o",
+  \u3070: "b a",
+  \u3073: "b i",
+  \u3076: "b u",
+  \u3079: "b e",
+  \u307C: "b o",
+  \u3071: "p a",
+  \u3074: "p i",
+  \u3077: "p u",
+  \u307A: "p e",
+  \u307D: "p o",
+  \u307E: "m a",
+  \u307F: "m i",
+  \u3080: "m u",
+  \u3081: "m e",
+  \u3082: "m o",
+  \u3084: "y a",
+  \u3086: "y u",
+  \u3088: "y o",
+  \u3089: "r a",
+  \u308A: "r i",
+  \u308B: "r u",
+  \u308C: "r e",
+  \u308D: "r o",
+  \u308F: "w a",
+  \u3090: "w i",
+  \u3091: "w e",
+  \u3092: "w o",
+  \u3094: "v u"
+};
+var SMALL = {
+  \u3041: "a",
+  \u3043: "i",
+  \u3045: "u",
+  \u3047: "e",
+  \u3049: "o",
+  \u3083: "a",
+  \u3085: "u",
+  \u3087: "o",
+  \u308E: "a"
+};
+var DIGRAPH = {
+  \u3057\u3083: "sh a",
+  \u3057\u3085: "sh u",
+  \u3057\u3087: "sh o",
+  \u3057\u3047: "sh e",
+  \u3058\u3083: "j a",
+  \u3058\u3085: "j u",
+  \u3058\u3087: "j o",
+  \u3058\u3047: "j e",
+  \u3061\u3083: "ch a",
+  \u3061\u3085: "ch u",
+  \u3061\u3087: "ch o",
+  \u3061\u3047: "ch e",
+  \u3062\u3083: "j a",
+  \u3062\u3085: "j u",
+  \u3062\u3087: "j o",
+  \u3064\u3041: "ts a",
+  \u3064\u3043: "ts i",
+  \u3064\u3047: "ts e",
+  \u3064\u3049: "ts o",
+  \u3066\u3043: "t i",
+  \u3066\u3083: "ty a",
+  \u3066\u3085: "ty u",
+  \u3066\u3087: "ty o",
+  \u3067\u3043: "d i",
+  \u3067\u3083: "dy a",
+  \u3067\u3085: "dy u",
+  \u3067\u3087: "dy o",
+  \u3068\u3045: "t u",
+  \u3069\u3045: "d u",
+  \u3075\u3041: "f a",
+  \u3075\u3043: "f i",
+  \u3075\u3047: "f e",
+  \u3075\u3049: "f o",
+  \u3075\u3085: "f u",
+  \u3094\u3041: "v a",
+  \u3094\u3043: "v i",
+  \u3094\u3047: "v e",
+  \u3094\u3049: "v o",
+  \u3094\u3085: "v u",
+  \u304F\u3041: "k a",
+  \u304F\u3043: "k i",
+  \u304F\u3047: "k e",
+  \u304F\u3049: "k o",
+  \u3050\u3041: "g a",
+  \u3050\u3043: "g i",
+  \u3050\u3047: "g e",
+  \u3050\u3049: "g o",
+  \u3059\u3043: "s i",
+  \u305A\u3043: "z i"
+};
+function toHiragana(s) {
+  let out = "";
+  for (const ch of s) {
+    const c = ch.codePointAt(0) ?? 0;
+    out += c >= 12449 && c <= 12534 ? String.fromCodePoint(c - 96) : ch;
+  }
+  return out;
+}
+function make(kana, pair) {
+  const sp = pair.indexOf(" ");
+  const consonant = pair.slice(0, sp);
+  const vowel = pair.slice(sp + 1);
+  return { kana, consonant, vowel, cls: CLASS_OF[consonant] ?? "vowel" };
+}
+function readSyllable(src, hira, i) {
+  const two = hira.slice(i, i + 2);
+  if (two.length === 2) {
+    const explicit = DIGRAPH[two];
+    if (explicit) return { syl: make(src.slice(i, i + 2), explicit), len: 2 };
+    const small = SMALL[two[1]];
+    const base = KANA[two[0]];
+    if (small && base) {
+      const sp = base.indexOf(" ");
+      const baseConsonant = base.slice(0, sp);
+      const baseVowel = base.slice(sp + 1);
+      if (baseVowel === "i" && CLASS_OF[`${baseConsonant}y`]) {
+        const pair = `${baseConsonant}y ${small}`;
+        return { syl: make(src.slice(i, i + 2), pair), len: 2 };
+      }
+      if (!baseConsonant) {
+        const glide = baseVowel === "i" ? "y" : baseVowel === "u" ? "w" : "";
+        const syl = make(src.slice(i, i + 2), `${glide} ${small}`);
+        return { syl: { ...syl, cls: "semivowel" }, len: 2 };
+      }
+      return {
+        syl: make(src.slice(i, i + 2), `${baseConsonant} ${small}`),
+        len: 2
+      };
+    }
+  }
+  const one = KANA[hira[i]];
+  if (one) return { syl: make(src.slice(i, i + 1), one), len: 1 };
+  return null;
+}
+function splitKana(text) {
+  const src = text.normalize("NFC");
+  const hira = toHiragana(src);
+  const out = [];
+  for (let i = 0; i < hira.length; ) {
+    const ch = hira[i];
+    if (ch === "\u3063" || ch === "\u30FC") {
+      i++;
+      continue;
+    }
+    const read = readSyllable(src, hira, i);
+    if (!read) return null;
+    out.push(read.syl);
+    i += read.len;
+  }
+  return out.length > 0 ? out : null;
+}
+
+// src/oto/generate.ts
+var REST = /[RrＲｒ]$|息$|吸$/;
+function transcribe(filename) {
+  let body = filename.replace(/\.[^.]+$/, "");
+  body = body.replace(/^[_\-\s]+/, "");
+  body = body.replace(/_[A-G][#b]?-?\d$/i, "");
+  body = body.trim();
+  if (!body) return null;
+  let mark = "";
+  const arrows = /[↑↓→]+$/.exec(body);
+  if (arrows) {
+    mark = arrows[0];
+    body = body.slice(0, -arrows[0].length);
+  }
+  let trailingRest = false;
+  if (REST.test(body)) {
+    trailingRest = true;
+    body = body.replace(REST, "");
+  }
+  const take = /[A-Za-z0-9'][A-Za-z0-9']?$/.exec(body);
+  if (take && body.length > take[0].length) {
+    mark = take[0] + mark;
+    body = body.slice(0, -take[0].length);
+  }
+  let prefix = "";
+  const head = /^[A-Za-z]{1,2}(?=[ぁ-ゟァ-ヿ])/.exec(body);
+  if (head) {
+    prefix = head[0];
+    body = body.slice(head[0].length);
+  }
+  if (!body) return null;
+  const syllables = splitKana(body);
+  if (!syllables) return null;
+  return { syllables, trailingRest, prefix, mark };
+}
+function suffixFromFolderName(folder) {
+  if (/^[A-G][#b]?-?\d$/.test(folder)) return `_${folder}`;
+  const tagged = /_([A-G][#b]?-?\d)(?![0-9A-Za-z])/.exec(folder);
+  return tagged ? `_${tagged[1]}` : "";
+}
+function aliasesFor(name, headAliases) {
+  return headAliases ? [name, `- ${name}`] : [name];
+}
+function generateOtoForFile(file, options = {}) {
+  const suffix = options.suffix ?? "";
+  const headAliases = options.headAliases ?? true;
+  const vowelJoinAliases = options.vowelJoinAliases ?? true;
+  const transcript = transcribe(file.name);
+  if (!transcript) {
+    return {
+      entries: [],
+      skipped: { wav: file.name, reason: "filename is not kana" },
+      style: null
+    };
+  }
+  let frames;
+  try {
+    frames = analyze(parseWav(file.data));
+  } catch (err) {
+    return {
+      entries: [],
+      skipped: {
+        wav: file.name,
+        reason: err instanceof Error ? err.message : String(err)
+      },
+      style: null
+    };
+  }
+  if (frames.n < 8) {
+    return {
+      entries: [],
+      skipped: { wav: file.name, reason: "too short to analyse" },
+      style: null
+    };
+  }
+  const { syllables, trailingRest, prefix, mark } = transcript;
+  if (syllables.length === 1 && !trailingRest) {
+    const syl = syllables[0];
+    const name = `${prefix}${syl.kana}${mark}${suffix}`;
+    const entries2 = estimateSolo(
+      file.name,
+      frames,
+      syl,
+      aliasesFor(name, headAliases)
+    );
+    if (vowelJoinAliases && (syl.cls === "vowel" || syl.cls === "nasalN")) {
+      const join2 = estimateVowelJoin(file.name, frames, syl, `* ${name}`);
+      if (join2) entries2.push(join2);
+    }
+    return { entries: entries2, skipped: null, style: "solo" };
+  }
+  const entries = estimateSequence(file.name, frames, syllables, {
+    suffix: `${mark}${suffix}`,
+    prefix,
+    trailingRest
+  });
+  if (entries.length === 0) {
+    return {
+      entries: [],
+      skipped: { wav: file.name, reason: "could not segment the phrase" },
+      style: null
+    };
+  }
+  return { entries, skipped: null, style: "sequence" };
+}
+function generateOto(files, options = {}) {
+  const entries = [];
+  const skipped = [];
+  let solo = 0;
+  let sequence = 0;
+  for (const file of files) {
+    const result = generateOtoForFile(file, options);
+    entries.push(...result.entries);
+    if (result.skipped) skipped.push(result.skipped);
+    if (result.style === "solo") solo++;
+    if (result.style === "sequence") sequence++;
+  }
+  return { entries, skipped, style: summarise(solo, sequence) };
+}
+function summarise(solo, sequence) {
+  if (solo && sequence) return "mixed";
+  if (sequence) return "sequence";
+  if (solo) return "solo";
+  return "empty";
+}
+
+// src/oto/write.ts
+var reverseTable = null;
+function shiftJisTable() {
+  if (reverseTable) return reverseTable;
+  const table = /* @__PURE__ */ new Map();
+  const decoder = new TextDecoder("shift_jis", { fatal: false });
+  for (let b = 0; b < 128; b++) {
+    table.set(decoder.decode(new Uint8Array([b])), b);
+  }
+  for (let b = 161; b <= 223; b++) {
+    table.set(decoder.decode(new Uint8Array([b])), b);
+  }
+  const pair = new Uint8Array(2);
+  for (let lead = 129; lead <= 252; lead++) {
+    if (lead > 159 && lead < 224) continue;
+    pair[0] = lead;
+    for (let trail = 64; trail <= 252; trail++) {
+      if (trail === 127) continue;
+      pair[1] = trail;
+      const ch = decoder.decode(pair);
+      if (ch.length !== 1 || ch === "\uFFFD") continue;
+      if (!table.has(ch)) table.set(ch, lead << 8 | trail);
+    }
+  }
+  reverseTable = table;
+  return table;
+}
+function encodeShiftJis(text) {
+  const table = shiftJisTable();
+  const out = [];
+  for (const ch of text) {
+    const code = table.get(ch);
+    if (code === void 0) out.push(63);
+    else if (code > 255) out.push(code >> 8, code & 255);
+    else out.push(code);
+  }
+  return new Uint8Array(out);
+}
+function num(v) {
+  const s = v.toFixed(3);
+  return s.replace(/\.?0+$/, "") || "0";
+}
+function formatOto(entries) {
+  return entries.map(
+    (e) => `${e.wav}=${e.alias},${num(e.offset)},${num(e.consonant)},${num(e.cutoff)},${num(e.pre)},${num(e.overlap)}`
+  ).join("\r\n").concat("\r\n");
+}
+function encodeOto(entries) {
+  return encodeShiftJis(formatOto(entries));
+}
+
+// src/oto/cli.ts
+function parseArgs(argv) {
+  let root = "";
+  let force = false;
+  let dryRun = false;
+  let quiet = false;
+  let suffix;
+  for (let i = 0; i < argv.length; i++) {
+    const a = argv[i];
+    if (a === "--force" || a === "-f") force = true;
+    else if (a === "--dry-run" || a === "-n") dryRun = true;
+    else if (a === "--quiet" || a === "-q") quiet = true;
+    else if (a === "--suffix") suffix = argv[++i] ?? "";
+    else if (a.startsWith("--suffix=")) suffix = a.slice(9);
+    else if (a.startsWith("-")) return null;
+    else if (!root) root = a;
+    else return null;
+  }
+  return root ? { root, force, dryRun, suffix, quiet } : null;
+}
+var USAGE = `Usage: koe-oto <voice-bank-dir> [options]
+
+Estimates UTAU oto.ini parameters from the recordings themselves and writes one
+oto.ini into every folder that holds WAV files.
+
+Options:
+  -f, --force        overwrite an existing oto.ini (a .bak copy is kept)
+  -n, --dry-run      report what would be written, write nothing
+      --suffix <s>   append <s> to every alias (default: the folder's note name)
+  -q, --quiet        only print the summary
+`;
+async function findWavFolders(root) {
+  const folders = /* @__PURE__ */ new Map();
+  const walk = async (dir) => {
+    const entries = await readdir(dir, { withFileTypes: true });
+    const wavs = entries.filter((e) => e.isFile() && /\.wav$/i.test(e.name)).map((e) => e.name).sort((a, b) => a.localeCompare(b, "ja"));
+    if (wavs.length > 0) folders.set(dir, wavs);
+    for (const e of entries) {
+      if (e.isDirectory()) await walk(join(dir, e.name));
+    }
+  };
+  await walk(root);
+  return folders;
+}
+function toArrayBuffer(buf) {
+  return buf.buffer.slice(
+    buf.byteOffset,
+    buf.byteOffset + buf.byteLength
+  );
+}
+async function exists(path) {
+  try {
+    await stat(path);
+    return true;
+  } catch {
+    return false;
+  }
+}
+async function main() {
+  const args = parseArgs(process.argv.slice(2));
+  if (!args) {
+    process.stderr.write(USAGE);
+    process.exit(1);
+  }
+  const root = resolve(args.root);
+  const folders = await findWavFolders(root);
+  if (folders.size === 0) {
+    process.stderr.write(`No WAV files found under "${root}"
+`);
+    process.exit(1);
+  }
+  const log = (s) => {
+    if (!args.quiet) process.stdout.write(s);
+  };
+  let totalEntries = 0;
+  let totalSkipped = 0;
+  let written = 0;
+  let blocked = 0;
+  for (const [dir, names] of folders) {
+    const label = relative(root, dir) || basename(root);
+    const files = [];
+    for (const name of names) {
+      files.push({
+        name,
+        data: toArrayBuffer(await readFile(join(dir, name)))
+      });
+    }
+    const suffix = args.suffix ?? suffixFromFolderName(basename(dir));
+    const result = generateOto(files, { suffix });
+    totalEntries += result.entries.length;
+    totalSkipped += result.skipped.length;
+    if (result.entries.length === 0) {
+      log(`  --  ${label}: no usable recordings (${names.length} wav)
+`);
+      continue;
+    }
+    const otoPath = join(dir, "oto.ini");
+    const already = await exists(otoPath);
+    const detail = `${result.entries.length} entries from ${names.length} wav [${result.style}]${result.skipped.length ? `, ${result.skipped.length} skipped` : ""}`;
+    if (args.dryRun) {
+      log(`  ..  ${label}: ${detail}${already ? " (would overwrite)" : ""}
+`);
+      continue;
+    }
+    if (already && !args.force) {
+      log(
+        `  !!  ${label}: oto.ini exists \u2014 rerun with --force to replace it
+`
+      );
+      blocked++;
+      continue;
+    }
+    if (already) {
+      await writeFile(`${otoPath}.bak`, await readFile(otoPath));
+    }
+    await writeFile(otoPath, encodeOto(result.entries));
+    written++;
+    log(`  ok  ${label}: ${detail}
+`);
+  }
+  process.stdout.write(
+    `${args.dryRun ? "Would write" : "Wrote"} ${args.dryRun ? folders.size : written} oto.ini \u2014 ${totalEntries} entries, ${totalSkipped} files skipped
+`
+  );
+  if (blocked > 0) {
+    process.stdout.write(
+      `${blocked} folder(s) already had an oto.ini; rerun with --force to replace them.
+`
+    );
+  }
+}
+main().catch((err) => {
+  process.stderr.write(`${err instanceof Error ? err.stack : err}
+`);
+  process.exit(1);
+});