sophontalk-services 0.0.2 → 0.0.4

package/dist-lib/sophontalk-services.cjs

@@ -1 +1,324 @@
-"use strict";Object.defineProperty(exports,Symbol.toStringTag,{value:"Module"});const l=16e3,_=25,u=10,w=80,m=4,M=.97,f=1e-10,k=20,F=!0,R=!0,h=25,g=1e3/h,p=t=>{if(t===128)return 4;if(t===96)return 3;throw new Error(`Unsupported crop size: ${t}. Only nano (96) and tiny (128) are supported.`)},O=(t,e,s)=>{if(e<=1)return{nextIndex:0,nextDirection:1};let r=t+s,a=s;return r>=e?(r=e-2,a=-1):r<0&&(r=1,a=1),{nextIndex:r,nextDirection:a}};async function L(t={}){const{jsonUrl:e="/complete_dataset.json",zipUrl:s="/processed_images.zip"}=t,[r,a]=await Promise.all([fetch(e),fetch(s)]);if(!r.ok)throw new Error(`无法加载 ${e}`);if(!a.ok)throw new Error(`无法加载 ${s}`);const o=await r.json(),c=await a.arrayBuffer();return{dataset:o,zipBuffer:c}}function S(t){return new Worker("./assets/feature.worker-Dx6moind.js",{type:"module",name:t?.name})}class P{worker=new S;async process(e){const s=e.getChannelData(0),r=e.numberOfChannels>1?e.getChannelData(1):void 0,a={leftChannel:s,rightChannel:r,sampleRate:e.sampleRate},o=[s.buffer];return r&&o.push(r.buffer),new Promise((c,d)=>{const i=n=>{n.data?.status==="success"?(this.worker.removeEventListener("message",i),this.worker.removeEventListener("error",E),c(n.data.payload)):n.data?.status==="error"&&(this.worker.removeEventListener("message",i),this.worker.removeEventListener("error",E),d(new Error(n.data.error||"特征提取失败")))},E=n=>{this.worker.removeEventListener("message",i),this.worker.removeEventListener("error",E),d(n)};this.worker.addEventListener("message",i),this.worker.addEventListener("error",E),this.worker.postMessage(a,o)})}dispose(){this.worker.terminate()}}function T(t){return new Worker("./assets/inference.worker-cwIlIomt.js",{type:"module",name:t?.name})}class A{worker;handlers={};constructor(e={}){this.handlers=e,this.worker=new T,this.worker.onmessage=this.handleMessage.bind(this)}handleMessage(e){const{type:s,payload:r}=e.data;switch(s){case"ready":this.handlers.onReady?.();break;case"start":this.handlers.onStart?.(r.totalFrames);break;case"frame":this.handlers.onFrame?.(r.frame,r.index);break;case"done":this.handlers.onDone?.();break;case"error":this.handlers.onError?.(r);break}}init(e){this.worker.postMessage({type:"init",modelPath:e.modelPath,dataset:e.dataset,zipBuffer:e.zipBuffer,blendingMask:e.blendingMask,wasmPaths:e.wasmPaths},[e.zipBuffer,e.blendingMask])}run(e){this.worker.postMessage({type:"run",audioFeatures:e.audioFeatures,audioDimensions:e.audioDimensions},[e.audioFeatures.buffer])}stop(){this.worker.postMessage({type:"stop"})}terminate(){this.worker.terminate()}}exports.DEFAULT_LOW_FREQ=k;exports.ENERGY_FLOOR=f;exports.FRAME_INTERVAL_MS=g;exports.FRAME_LENGTH_MS=_;exports.FRAME_SHIFT_MS=u;exports.FeatureExtractor=P;exports.InferenceEngine=A;exports.NUM_MEL_BINS=w;exports.NUM_SEQUENCE_FRAMES=m;exports.PREEMPH_COEFF=M;exports.REMOVE_DC_OFFSET=F;exports.ROUND_TO_POWER_OF_TWO=R;exports.SAMPLE_RATE=l;exports.TARGET_FPS=h;exports.calculatePingPongState=O;exports.getBorder=p;exports.loadDataset=L;
+"use strict";Object.defineProperty(exports,Symbol.toStringTag,{value:"Module"});const w=16e3,g=25,_=10,y=80,E=4,M=.97,R=1e-10,b=20,F=!0,L=!0,d=25,I=1e3/d,k=a=>{if(a===128)return 4;if(a===96)return 3;throw new Error(`Unsupported crop size: ${a}. Only nano (96) and tiny (128) are supported.`)},v=(a,n,e)=>{if(n<=1)return{nextIndex:0,nextDirection:1};let t=a+e,r=e;return t>=n?(t=n-2,r=-1):t<0&&(t=1,r=1),{nextIndex:t,nextDirection:r}};async function U(a={}){const{jsonUrl:n="/complete_dataset.json",zipUrl:e="/processed_images.zip"}=a,[t,r]=await Promise.all([fetch(n),fetch(e)]);if(!t.ok)throw new Error(`无法加载 ${n}`);if(!r.ok)throw new Error(`无法加载 ${e}`);const o=await t.json(),c=await r.arrayBuffer();return{dataset:o,zipBuffer:c}}const h=`import x from "fft.js";
+import { create as O, ConverterType as H } from "@alexanderolsen/libsamplerate-js";
+const y = 16e3, U = 25, D = 10, u = 80, S = 4, W = 0.97, L = 1e-10, b = 20, h = Math.round(U * y / 1e3), C = Math.round(D * y / 1e3), k = (e) => {
+  if (e <= 0) return 1;
+  let t = 1;
+  for (; t < e; ) t *= 2;
+  return t;
+}, P = k(h), N = new x(P), v = new Float32Array(h), G = () => {
+  if (h <= 1) return;
+  const e = 2 * Math.PI / (h - 1);
+  for (let t = 0; t < h; t++)
+    v[t] = 0.5 - 0.5 * Math.cos(t * e);
+};
+G();
+function Q(e, t, o, a = 20, s) {
+  s = s || o / 2;
+  const i = (n) => 700 * (Math.exp(n / 1127) - 1), r = (n) => 1127 * Math.log(1 + n / 700), l = r(a), A = r(s), w = new Float32Array(e + 2);
+  for (let n = 0; n < e + 2; n++)
+    w[n] = l + (A - l) * n / (e + 1);
+  const _ = w.map(i).map((n) => n * t / o), f = Math.floor(t / 2) + 1, m = [];
+  for (let n = 1; n <= e; n++) {
+    const c = _[n - 1], F = _[n], M = _[n + 1], I = Math.ceil(c), T = Math.floor(M), E = Math.max(0, I), p = Math.min(f - 1, T), B = p - E + 1;
+    if (B <= 0) {
+      m.push({ startBin: 0, values: new Float32Array(0) });
+      continue;
+    }
+    const R = new Float32Array(B);
+    for (let d = E; d <= p; d++) {
+      let g = 0;
+      d <= F ? g = (d - c) / (F - c) : g = (M - d) / (M - F), R[d - E] = g;
+    }
+    m.push({ startBin: E, values: R });
+  }
+  return m;
+}
+const q = Q(
+  u,
+  P,
+  y,
+  b
+);
+function K(e) {
+  for (let t = 0; t < e.length; t++)
+    e[t] *= v[t];
+}
+function Y(e, t) {
+  for (let o = e.length - 1; o > 0; o--)
+    e[o] -= t * e[o - 1];
+  e[0] -= t * e[0];
+}
+function V(e) {
+  const t = e.length;
+  if (t === 0) return;
+  let o = 0;
+  for (let s = 0; s < t; s++) o += e[s];
+  const a = o / t;
+  for (let s = 0; s < t; s++) e[s] -= a;
+}
+function Z(e, t) {
+  const o = e.length;
+  for (let a = 0; a < o; a++) {
+    const s = t[a * 2], i = t[a * 2 + 1];
+    e[a] = s * s + i * i;
+  }
+}
+function j(e) {
+  const t = e.length;
+  if (t < 2) return;
+  let o = 0;
+  for (let r = 0; r < t; r++) o += e[r];
+  const a = o / t;
+  let s = 0;
+  for (let r = 0; r < t; r++) {
+    const l = e[r] - a;
+    s += l * l;
+  }
+  const i = Math.sqrt(s / (t - 1));
+  if (i > 1e-8) {
+    const r = 1 / i;
+    for (let l = 0; l < t; l++)
+      e[l] = (e[l] - a) * r;
+  } else
+    for (let r = 0; r < t; r++)
+      e[r] -= a;
+}
+async function z(e) {
+  let t = e.leftChannel;
+  if (e.rightChannel) {
+    const f = e.rightChannel, m = new Float32Array(t.length);
+    for (let n = 0; n < t.length; n++)
+      m[n] = (t[n] + f[n]) * 0.5;
+    t = m;
+  }
+  if (e.sampleRate !== y) {
+    const f = await O(1, e.sampleRate, y, {
+      converterType: H.SRC_SINC_MEDIUM_QUALITY
+    });
+    t = f.simple(t), f.destroy();
+  }
+  const o = Math.floor((t.length - h) / C) + 1;
+  if (o <= 0)
+    throw new Error("特征提取失败：音频时长过短。");
+  const a = Math.ceil(o / S) * S, s = a * u, i = new Float32Array(s), r = new Float32Array(h), l = N.createComplexArray(), A = N.createComplexArray(), w = new Float32Array(Math.floor(P / 2) + 1);
+  for (let f = 0; f < o; f++) {
+    const m = f * C;
+    r.set(t.subarray(m, m + h)), V(r), Y(r, W), K(r), l.fill(0);
+    for (let c = 0; c < h; c++)
+      l[c * 2] = r[c];
+    N.transform(A, l), Z(w, A);
+    const n = f * u;
+    for (let c = 0; c < u; c++) {
+      const F = q[c];
+      let M = 0;
+      const I = F.values, T = F.startBin, E = I.length;
+      for (let p = 0; p < E; p++)
+        M += w[T + p] * I[p];
+      M = M < L ? L : M, i[n + c] = Math.log(M);
+    }
+  }
+  if (a > o) {
+    const f = (o - 1) * u, m = f + u, n = i.subarray(f, m);
+    for (let c = o; c < a; c++)
+      i.set(n, c * u);
+  }
+  j(i);
+  const _ = a / S;
+  return {
+    features: i,
+    dimensions: [_, S, u]
+  };
+}
+self.addEventListener(
+  "message",
+  async (e) => {
+    try {
+      const t = e.data, o = await z(t);
+      self.postMessage(
+        { status: "success", payload: o },
+        { transfer: [o.features.buffer] }
+      );
+    } catch (t) {
+      self.postMessage({ status: "error", error: t.message });
+    }
+  }
+);
+`,u=typeof self<"u"&&self.Blob&&new Blob(["URL.revokeObjectURL(import.meta.url);",h],{type:"text/javascript;charset=utf-8"});function C(a){let n;try{if(n=u&&(self.URL||self.webkitURL).createObjectURL(u),!n)throw"";const e=new Worker(n,{type:"module",name:a?.name});return e.addEventListener("error",()=>{(self.URL||self.webkitURL).revokeObjectURL(n)}),e}catch{return new Worker("data:text/javascript;charset=utf-8,"+encodeURIComponent(h),{type:"module",name:a?.name})}}class x{worker=new C;async process(n){const e=n.getChannelData(0),t=n.numberOfChannels>1?n.getChannelData(1):void 0,r={leftChannel:e,rightChannel:t,sampleRate:n.sampleRate},o=[e.buffer];return t&&o.push(t.buffer),new Promise((c,f)=>{const l=s=>{s.data?.status==="success"?(this.worker.removeEventListener("message",l),this.worker.removeEventListener("error",i),c(s.data.payload)):s.data?.status==="error"&&(this.worker.removeEventListener("message",l),this.worker.removeEventListener("error",i),f(new Error(s.data.error||"特征提取失败")))},i=s=>{this.worker.removeEventListener("message",l),this.worker.removeEventListener("error",i),f(s)};this.worker.addEventListener("message",l),this.worker.addEventListener("error",i),this.worker.postMessage(r,o)})}dispose(){this.worker.terminate()}}const p=`import { env as k, InferenceSession as $, Tensor as W } from "onnxruntime-web";
+import q from "jszip";
+const V = 80, G = 4, H = 25, j = 1e3 / H, J = (t) => {
+  if (t === 128) return 4;
+  if (t === 96) return 3;
+  throw new Error(\`Unsupported crop size: \${t}. Only nano (96) and tiny (128) are supported.\`);
+}, Q = (t, e, n) => {
+  if (e <= 1) return { nextIndex: 0, nextDirection: 1 };
+  let a = t + n, s = n;
+  return a >= e ? (a = e - 2, s = -1) : a < 0 && (a = 1, s = 1), { nextIndex: a, nextDirection: s };
+};
+k.wasm.numThreads = 1;
+const Z = () => typeof navigator < "u" && !!navigator.gpu, F = 32;
+let _ = null, R = null, x = null, L = null;
+const w = {
+  tensors: /* @__PURE__ */ new Map(),
+  faceImages: /* @__PURE__ */ new Map(),
+  fullImages: /* @__PURE__ */ new Map()
+};
+let S = 3, b = 96, O = 0, z = 0, C = null, A = null, y = null, E = null, g = null, I = null, D = null, M = null, P = !1, v = "", p = null;
+const T = (t) => {
+  self.postMessage({ type: "error", payload: t });
+}, K = (t) => {
+  const e = w.tensors.get(t);
+  if (!e) throw new Error(\`Cache miss: \${t}\`);
+  const n = b - 2 * S;
+  return new W("float32", e, [1, 6, n, n]);
+}, X = (t) => {
+  const e = w.faceImages.get(t);
+  if (!e) throw new Error(\`Cache miss: \${t}\`);
+  return e;
+}, Y = async (t) => {
+  const e = w.fullImages.get(t);
+  if (!e) throw new Error(\`Cache miss: \${t}\`);
+  return await createImageBitmap(e);
+}, ee = (t) => {
+  const [e, n, a, s] = t.dims, u = t.data;
+  (!y || y.width !== s || y.height !== a) && (C.width = s, C.height = a, y = A.createImageData(s, a));
+  const d = a * s, l = new Uint32Array(y.data.buffer), o = d, i = d * 2, c = n === 3;
+  for (let r = 0; r < d; r++) {
+    let m = (c ? u[r + i] : u[r]) * 255, f = (c ? u[r + o] : u[r]) * 255, h = u[r] * 255;
+    m = m < 0 ? 0 : m > 255 ? 255 : m, f = f < 0 ? 0 : f > 255 ? 255 : f, h = h < 0 ? 0 : h > 255 ? 255 : h, l[r] = 255 << 24 | (h & 255) << 16 | (f & 255) << 8 | m & 255;
+  }
+  A.putImageData(y, 0, 0);
+}, te = async (t, e, n) => {
+  ee(t);
+  const a = X(e.face_image);
+  g.globalCompositeOperation = "source-over", g.drawImage(a, 0, 0), g.drawImage(C, S, S), g.globalCompositeOperation = "destination-in", g.drawImage(L, 0, 0, b, b), g.globalCompositeOperation = "destination-over", g.drawImage(a, 0, 0), g.globalCompositeOperation = "source-over", D.drawImage(n, 0, 0);
+  const s = e.crop_info.ymax - e.crop_info.ymin;
+  return D.drawImage(
+    E,
+    e.crop_info.xmin,
+    e.crop_info.ymin,
+    e.crop_info.width,
+    s
+  ), await createImageBitmap(I);
+}, ae = (t, e, n) => {
+  const [a, s, u] = e, d = s * u, l = F * d;
+  (!M || M.length !== l) && (M = new Float32Array(l));
+  for (let o = 0; o < F; o++) {
+    const c = Math.min(
+      Math.max(n - F / 2 + o, 0),
+      a - 1
+    ) * d, r = o * d;
+    M.set(
+      t.subarray(c, c + d),
+      r
+    );
+  }
+  return M;
+}, ne = (t, e, n) => {
+  C || (C = new OffscreenCanvas(1, 1), A = C.getContext("2d", {
+    willReadFrequently: !0
+  })), (!E || E.width !== n) && (E = new OffscreenCanvas(n, n), g = E.getContext("2d", {
+    willReadFrequently: !0
+  })), (!I || I.width !== t || I.height !== e) && (I = new OffscreenCanvas(t, e), D = I.getContext("2d", {
+    willReadFrequently: !0
+  }));
+}, se = async (t) => {
+  try {
+    if (t.wasmPaths && (k.wasm.wasmPaths = t.wasmPaths), !t.dataset) throw new Error("Missing dataset");
+    const e = t.dataset;
+    R = e, b = e.dataset_info.config.crop_size, S = J(b), O = e.dataset_info.source_image_dimensions?.width || 0, z = e.dataset_info.source_image_dimensions?.height || 0, ne(O, z, b), x = await q.loadAsync(t.zipBuffer), t.zipBuffer = null, L = t.blendingMask;
+    const n = e.images.map(async (l) => {
+      if (!w.tensors.has(l.tensor_file)) {
+        const o = x.file(l.tensor_file);
+        if (o) {
+          const i = await o.async("arraybuffer");
+          w.tensors.set(l.tensor_file, new Float32Array(i));
+        }
+      }
+      if (!w.faceImages.has(l.face_image)) {
+        const o = x.file(l.face_image);
+        if (o) {
+          const i = await o.async("blob"), c = await createImageBitmap(i);
+          w.faceImages.set(l.face_image, c);
+        }
+      }
+      if (!w.fullImages.has(l.full_image)) {
+        const o = x.file(l.full_image);
+        if (o) {
+          const i = await o.async("blob");
+          w.fullImages.set(l.full_image, i);
+        }
+      }
+    });
+    await Promise.all(n), x = null;
+    const a = [];
+    Z() && a.push("webgpu"), a.push("wasm");
+    const s = new URL(t.modelPath, self.location.href).toString(), u = new URL(s).pathname.split("/").pop() || "model.onnx", d = new URL(\`\${u}.data\`, s).toString();
+    _ = await $.create(s, {
+      executionProviders: a,
+      graphOptimizationLevel: "all",
+      enableMemPattern: !0,
+      enableCpuMemArena: !0,
+      externalData: [{ data: d, path: \`\${u}.data\` }]
+    }), self.postMessage({ type: "ready" });
+  } catch (e) {
+    T(e?.message || "Init Error");
+  }
+}, oe = async (t) => {
+  if (!_ || !R) {
+    T("Not initialized");
+    return;
+  }
+  P = !1;
+  const e = t.audioDimensions[0];
+  self.postMessage({
+    type: "start",
+    payload: { totalFrames: e, frameInterval: j }
+  });
+  let n = 0, a = 1;
+  const s = {}, u = _.inputNames[0], d = _.inputNames[1], l = _.outputNames[0];
+  v = "", p && (p.close(), p = null);
+  for (let o = 0; o < e && !P; o++) {
+    let i = null, c = null, r = null, m = null;
+    try {
+      const f = R.images[n];
+      i = K(f.tensor_file), f.full_image !== v && (p && p.close(), p = await Y(f.full_image), v = f.full_image);
+      const h = ae(
+        t.audioFeatures,
+        t.audioDimensions,
+        o
+      );
+      c = new W("float32", h, [
+        1,
+        F,
+        G,
+        V
+      ]), s[u] = i, s[d] = c, r = await _.run(s);
+      const N = r[l];
+      if (!N)
+        throw new Error("Missing inference output tensor");
+      m = N;
+      const B = await te(N, f, p);
+      self.postMessage(
+        { type: "frame", payload: { frame: B, index: o } },
+        [B]
+      );
+      const U = Q(
+        n,
+        R.images.length,
+        a
+      );
+      n = U.nextIndex, a = U.nextDirection;
+    } catch (f) {
+      T(f?.message || "Run Error");
+      break;
+    } finally {
+      i && i.dispose(), c && c.dispose(), m && m.dispose(), i = null, c = null, m = null, r = null;
+    }
+  }
+  p && (p.close(), p = null), v = "", self.postMessage({ type: "done" });
+};
+self.onmessage = (t) => {
+  const e = t.data;
+  e.type === "init" ? se(e) : e.type === "run" ? oe(e) : e.type === "stop" && (P = !0);
+};
+`,m=typeof self<"u"&&self.Blob&&new Blob(["URL.revokeObjectURL(import.meta.url);",p],{type:"text/javascript;charset=utf-8"});function O(a){let n;try{if(n=m&&(self.URL||self.webkitURL).createObjectURL(m),!n)throw"";const e=new Worker(n,{type:"module",name:a?.name});return e.addEventListener("error",()=>{(self.URL||self.webkitURL).revokeObjectURL(n)}),e}catch{return new Worker("data:text/javascript;charset=utf-8,"+encodeURIComponent(p),{type:"module",name:a?.name})}}class A{worker;handlers={};constructor(n={}){this.handlers=n,this.worker=new O,this.worker.onmessage=this.handleMessage.bind(this)}handleMessage(n){const{type:e,payload:t}=n.data;switch(e){case"ready":this.handlers.onReady?.();break;case"start":this.handlers.onStart?.(t.totalFrames);break;case"frame":this.handlers.onFrame?.(t.frame,t.index);break;case"done":this.handlers.onDone?.();break;case"error":this.handlers.onError?.(t);break}}init(n){this.worker.postMessage({type:"init",modelPath:n.modelPath,dataset:n.dataset,zipBuffer:n.zipBuffer,blendingMask:n.blendingMask,wasmPaths:n.wasmPaths},[n.zipBuffer,n.blendingMask])}run(n){this.worker.postMessage({type:"run",audioFeatures:n.audioFeatures,audioDimensions:n.audioDimensions},[n.audioFeatures.buffer])}stop(){this.worker.postMessage({type:"stop"})}terminate(){this.worker.terminate()}}exports.DEFAULT_LOW_FREQ=b;exports.ENERGY_FLOOR=R;exports.FRAME_INTERVAL_MS=I;exports.FRAME_LENGTH_MS=g;exports.FRAME_SHIFT_MS=_;exports.FeatureExtractor=x;exports.InferenceEngine=A;exports.NUM_MEL_BINS=y;exports.NUM_SEQUENCE_FRAMES=E;exports.PREEMPH_COEFF=M;exports.REMOVE_DC_OFFSET=F;exports.ROUND_TO_POWER_OF_TWO=L;exports.SAMPLE_RATE=w;exports.TARGET_FPS=d;exports.calculatePingPongState=v;exports.getBorder=k;exports.loadDataset=U;

package/dist-lib/sophontalk-services.js

@@ -1,111 +1,460 @@
-const w = 16e3, m = 25, u = 10, k = 80, f = 4, p = 0.97, _ = 1e-10, g = 20, M = !0, R = !0, F = 25, L = 40, y = (t) => {
-  if (t === 128) return 4;
-  if (t === 96) return 3;
-  throw new Error(`Unsupported crop size: ${t}. Only nano (96) and tiny (128) are supported.`);
-}, P = (t, e, s) => {
-  if (e <= 1) return { nextIndex: 0, nextDirection: 1 };
-  let r = t + s, n = s;
-  return r >= e ? (r = e - 2, n = -1) : r < 0 && (r = 1, n = 1), { nextIndex: r, nextDirection: n };
+const g = 16e3, y = 25, _ = 10, M = 80, E = 4, b = 0.97, R = 1e-10, k = 20, I = !0, v = !0, L = 25, x = 40, F = (a) => {
+  if (a === 128) return 4;
+  if (a === 96) return 3;
+  throw new Error(`Unsupported crop size: ${a}. Only nano (96) and tiny (128) are supported.`);
+}, C = (a, n, e) => {
+  if (n <= 1) return { nextIndex: 0, nextDirection: 1 };
+  let t = a + e, r = e;
+  return t >= n ? (t = n - 2, r = -1) : t < 0 && (t = 1, r = 1), { nextIndex: t, nextDirection: r };
 };
-async function b(t = {}) {
+async function U(a = {}) {
   const {
-    jsonUrl: e = "/complete_dataset.json",
-    zipUrl: s = "/processed_images.zip"
-  } = t, [r, n] = await Promise.all([
-    fetch(e),
-    fetch(s)
+    jsonUrl: n = "/complete_dataset.json",
+    zipUrl: e = "/processed_images.zip"
+  } = a, [t, r] = await Promise.all([
+    fetch(n),
+    fetch(e)
   ]);
+  if (!t.ok) throw new Error(`无法加载 ${n}`);
   if (!r.ok) throw new Error(`无法加载 ${e}`);
-  if (!n.ok) throw new Error(`无法加载 ${s}`);
-  const o = await r.json(), h = await n.arrayBuffer();
-  return { dataset: o, zipBuffer: h };
+  const o = await t.json(), c = await r.arrayBuffer();
+  return { dataset: o, zipBuffer: c };
 }
-function l(t) {
-  return new Worker(
-    "./assets/feature.worker-Dx6moind.js",
-    {
-      type: "module",
-      name: t?.name
+const h = `import x from "fft.js";
+import { create as O, ConverterType as H } from "@alexanderolsen/libsamplerate-js";
+const y = 16e3, U = 25, D = 10, u = 80, S = 4, W = 0.97, L = 1e-10, b = 20, h = Math.round(U * y / 1e3), C = Math.round(D * y / 1e3), k = (e) => {
+  if (e <= 0) return 1;
+  let t = 1;
+  for (; t < e; ) t *= 2;
+  return t;
+}, P = k(h), N = new x(P), v = new Float32Array(h), G = () => {
+  if (h <= 1) return;
+  const e = 2 * Math.PI / (h - 1);
+  for (let t = 0; t < h; t++)
+    v[t] = 0.5 - 0.5 * Math.cos(t * e);
+};
+G();
+function Q(e, t, o, a = 20, s) {
+  s = s || o / 2;
+  const i = (n) => 700 * (Math.exp(n / 1127) - 1), r = (n) => 1127 * Math.log(1 + n / 700), l = r(a), A = r(s), w = new Float32Array(e + 2);
+  for (let n = 0; n < e + 2; n++)
+    w[n] = l + (A - l) * n / (e + 1);
+  const _ = w.map(i).map((n) => n * t / o), f = Math.floor(t / 2) + 1, m = [];
+  for (let n = 1; n <= e; n++) {
+    const c = _[n - 1], F = _[n], M = _[n + 1], I = Math.ceil(c), T = Math.floor(M), E = Math.max(0, I), p = Math.min(f - 1, T), B = p - E + 1;
+    if (B <= 0) {
+      m.push({ startBin: 0, values: new Float32Array(0) });
+      continue;
+    }
+    const R = new Float32Array(B);
+    for (let d = E; d <= p; d++) {
+      let g = 0;
+      d <= F ? g = (d - c) / (F - c) : g = (M - d) / (M - F), R[d - E] = g;
     }
-  );
+    m.push({ startBin: E, values: R });
+  }
+  return m;
 }
-class O {
-  worker = new l();
-  async process(e) {
-    const s = e.getChannelData(0), r = e.numberOfChannels > 1 ? e.getChannelData(1) : void 0, n = {
-      leftChannel: s,
-      rightChannel: r,
-      sampleRate: e.sampleRate
-    }, o = [s.buffer];
-    return r && o.push(r.buffer), new Promise((h, d) => {
-      const i = (a) => {
-        a.data?.status === "success" ? (this.worker.removeEventListener("message", i), this.worker.removeEventListener("error", c), h(a.data.payload)) : a.data?.status === "error" && (this.worker.removeEventListener("message", i), this.worker.removeEventListener("error", c), d(new Error(a.data.error || "特征提取失败")));
-      }, c = (a) => {
-        this.worker.removeEventListener("message", i), this.worker.removeEventListener("error", c), d(a);
+const q = Q(
+  u,
+  P,
+  y,
+  b
+);
+function K(e) {
+  for (let t = 0; t < e.length; t++)
+    e[t] *= v[t];
+}
+function Y(e, t) {
+  for (let o = e.length - 1; o > 0; o--)
+    e[o] -= t * e[o - 1];
+  e[0] -= t * e[0];
+}
+function V(e) {
+  const t = e.length;
+  if (t === 0) return;
+  let o = 0;
+  for (let s = 0; s < t; s++) o += e[s];
+  const a = o / t;
+  for (let s = 0; s < t; s++) e[s] -= a;
+}
+function Z(e, t) {
+  const o = e.length;
+  for (let a = 0; a < o; a++) {
+    const s = t[a * 2], i = t[a * 2 + 1];
+    e[a] = s * s + i * i;
+  }
+}
+function j(e) {
+  const t = e.length;
+  if (t < 2) return;
+  let o = 0;
+  for (let r = 0; r < t; r++) o += e[r];
+  const a = o / t;
+  let s = 0;
+  for (let r = 0; r < t; r++) {
+    const l = e[r] - a;
+    s += l * l;
+  }
+  const i = Math.sqrt(s / (t - 1));
+  if (i > 1e-8) {
+    const r = 1 / i;
+    for (let l = 0; l < t; l++)
+      e[l] = (e[l] - a) * r;
+  } else
+    for (let r = 0; r < t; r++)
+      e[r] -= a;
+}
+async function z(e) {
+  let t = e.leftChannel;
+  if (e.rightChannel) {
+    const f = e.rightChannel, m = new Float32Array(t.length);
+    for (let n = 0; n < t.length; n++)
+      m[n] = (t[n] + f[n]) * 0.5;
+    t = m;
+  }
+  if (e.sampleRate !== y) {
+    const f = await O(1, e.sampleRate, y, {
+      converterType: H.SRC_SINC_MEDIUM_QUALITY
+    });
+    t = f.simple(t), f.destroy();
+  }
+  const o = Math.floor((t.length - h) / C) + 1;
+  if (o <= 0)
+    throw new Error("特征提取失败：音频时长过短。");
+  const a = Math.ceil(o / S) * S, s = a * u, i = new Float32Array(s), r = new Float32Array(h), l = N.createComplexArray(), A = N.createComplexArray(), w = new Float32Array(Math.floor(P / 2) + 1);
+  for (let f = 0; f < o; f++) {
+    const m = f * C;
+    r.set(t.subarray(m, m + h)), V(r), Y(r, W), K(r), l.fill(0);
+    for (let c = 0; c < h; c++)
+      l[c * 2] = r[c];
+    N.transform(A, l), Z(w, A);
+    const n = f * u;
+    for (let c = 0; c < u; c++) {
+      const F = q[c];
+      let M = 0;
+      const I = F.values, T = F.startBin, E = I.length;
+      for (let p = 0; p < E; p++)
+        M += w[T + p] * I[p];
+      M = M < L ? L : M, i[n + c] = Math.log(M);
+    }
+  }
+  if (a > o) {
+    const f = (o - 1) * u, m = f + u, n = i.subarray(f, m);
+    for (let c = o; c < a; c++)
+      i.set(n, c * u);
+  }
+  j(i);
+  const _ = a / S;
+  return {
+    features: i,
+    dimensions: [_, S, u]
+  };
+}
+self.addEventListener(
+  "message",
+  async (e) => {
+    try {
+      const t = e.data, o = await z(t);
+      self.postMessage(
+        { status: "success", payload: o },
+        { transfer: [o.features.buffer] }
+      );
+    } catch (t) {
+      self.postMessage({ status: "error", error: t.message });
+    }
+  }
+);
+`, u = typeof self < "u" && self.Blob && new Blob(["URL.revokeObjectURL(import.meta.url);", h], { type: "text/javascript;charset=utf-8" });
+function p(a) {
+  let n;
+  try {
+    if (n = u && (self.URL || self.webkitURL).createObjectURL(u), !n) throw "";
+    const e = new Worker(n, {
+      type: "module",
+      name: a?.name
+    });
+    return e.addEventListener("error", () => {
+      (self.URL || self.webkitURL).revokeObjectURL(n);
+    }), e;
+  } catch {
+    return new Worker(
+      "data:text/javascript;charset=utf-8," + encodeURIComponent(h),
+      {
+        type: "module",
+        name: a?.name
+      }
+    );
+  }
+}
+class A {
+  worker = new p();
+  async process(n) {
+    const e = n.getChannelData(0), t = n.numberOfChannels > 1 ? n.getChannelData(1) : void 0, r = {
+      leftChannel: e,
+      rightChannel: t,
+      sampleRate: n.sampleRate
+    }, o = [e.buffer];
+    return t && o.push(t.buffer), new Promise((c, f) => {
+      const l = (s) => {
+        s.data?.status === "success" ? (this.worker.removeEventListener("message", l), this.worker.removeEventListener("error", i), c(s.data.payload)) : s.data?.status === "error" && (this.worker.removeEventListener("message", l), this.worker.removeEventListener("error", i), f(new Error(s.data.error || "特征提取失败")));
+      }, i = (s) => {
+        this.worker.removeEventListener("message", l), this.worker.removeEventListener("error", i), f(s);
       };
-      this.worker.addEventListener("message", i), this.worker.addEventListener("error", c), this.worker.postMessage(n, o);
+      this.worker.addEventListener("message", l), this.worker.addEventListener("error", i), this.worker.postMessage(r, o);
     });
   }
   dispose() {
     this.worker.terminate();
   }
 }
-function E(t) {
-  return new Worker(
-    "./assets/inference.worker-cwIlIomt.js",
-    {
-      type: "module",
-      name: t?.name
+const d = `import { env as k, InferenceSession as $, Tensor as W } from "onnxruntime-web";
+import q from "jszip";
+const V = 80, G = 4, H = 25, j = 1e3 / H, J = (t) => {
+  if (t === 128) return 4;
+  if (t === 96) return 3;
+  throw new Error(\`Unsupported crop size: \${t}. Only nano (96) and tiny (128) are supported.\`);
+}, Q = (t, e, n) => {
+  if (e <= 1) return { nextIndex: 0, nextDirection: 1 };
+  let a = t + n, s = n;
+  return a >= e ? (a = e - 2, s = -1) : a < 0 && (a = 1, s = 1), { nextIndex: a, nextDirection: s };
+};
+k.wasm.numThreads = 1;
+const Z = () => typeof navigator < "u" && !!navigator.gpu, F = 32;
+let _ = null, R = null, x = null, L = null;
+const w = {
+  tensors: /* @__PURE__ */ new Map(),
+  faceImages: /* @__PURE__ */ new Map(),
+  fullImages: /* @__PURE__ */ new Map()
+};
+let S = 3, b = 96, O = 0, z = 0, C = null, A = null, y = null, E = null, g = null, I = null, D = null, M = null, P = !1, v = "", p = null;
+const T = (t) => {
+  self.postMessage({ type: "error", payload: t });
+}, K = (t) => {
+  const e = w.tensors.get(t);
+  if (!e) throw new Error(\`Cache miss: \${t}\`);
+  const n = b - 2 * S;
+  return new W("float32", e, [1, 6, n, n]);
+}, X = (t) => {
+  const e = w.faceImages.get(t);
+  if (!e) throw new Error(\`Cache miss: \${t}\`);
+  return e;
+}, Y = async (t) => {
+  const e = w.fullImages.get(t);
+  if (!e) throw new Error(\`Cache miss: \${t}\`);
+  return await createImageBitmap(e);
+}, ee = (t) => {
+  const [e, n, a, s] = t.dims, u = t.data;
+  (!y || y.width !== s || y.height !== a) && (C.width = s, C.height = a, y = A.createImageData(s, a));
+  const d = a * s, l = new Uint32Array(y.data.buffer), o = d, i = d * 2, c = n === 3;
+  for (let r = 0; r < d; r++) {
+    let m = (c ? u[r + i] : u[r]) * 255, f = (c ? u[r + o] : u[r]) * 255, h = u[r] * 255;
+    m = m < 0 ? 0 : m > 255 ? 255 : m, f = f < 0 ? 0 : f > 255 ? 255 : f, h = h < 0 ? 0 : h > 255 ? 255 : h, l[r] = 255 << 24 | (h & 255) << 16 | (f & 255) << 8 | m & 255;
+  }
+  A.putImageData(y, 0, 0);
+}, te = async (t, e, n) => {
+  ee(t);
+  const a = X(e.face_image);
+  g.globalCompositeOperation = "source-over", g.drawImage(a, 0, 0), g.drawImage(C, S, S), g.globalCompositeOperation = "destination-in", g.drawImage(L, 0, 0, b, b), g.globalCompositeOperation = "destination-over", g.drawImage(a, 0, 0), g.globalCompositeOperation = "source-over", D.drawImage(n, 0, 0);
+  const s = e.crop_info.ymax - e.crop_info.ymin;
+  return D.drawImage(
+    E,
+    e.crop_info.xmin,
+    e.crop_info.ymin,
+    e.crop_info.width,
+    s
+  ), await createImageBitmap(I);
+}, ae = (t, e, n) => {
+  const [a, s, u] = e, d = s * u, l = F * d;
+  (!M || M.length !== l) && (M = new Float32Array(l));
+  for (let o = 0; o < F; o++) {
+    const c = Math.min(
+      Math.max(n - F / 2 + o, 0),
+      a - 1
+    ) * d, r = o * d;
+    M.set(
+      t.subarray(c, c + d),
+      r
+    );
+  }
+  return M;
+}, ne = (t, e, n) => {
+  C || (C = new OffscreenCanvas(1, 1), A = C.getContext("2d", {
+    willReadFrequently: !0
+  })), (!E || E.width !== n) && (E = new OffscreenCanvas(n, n), g = E.getContext("2d", {
+    willReadFrequently: !0
+  })), (!I || I.width !== t || I.height !== e) && (I = new OffscreenCanvas(t, e), D = I.getContext("2d", {
+    willReadFrequently: !0
+  }));
+}, se = async (t) => {
+  try {
+    if (t.wasmPaths && (k.wasm.wasmPaths = t.wasmPaths), !t.dataset) throw new Error("Missing dataset");
+    const e = t.dataset;
+    R = e, b = e.dataset_info.config.crop_size, S = J(b), O = e.dataset_info.source_image_dimensions?.width || 0, z = e.dataset_info.source_image_dimensions?.height || 0, ne(O, z, b), x = await q.loadAsync(t.zipBuffer), t.zipBuffer = null, L = t.blendingMask;
+    const n = e.images.map(async (l) => {
+      if (!w.tensors.has(l.tensor_file)) {
+        const o = x.file(l.tensor_file);
+        if (o) {
+          const i = await o.async("arraybuffer");
+          w.tensors.set(l.tensor_file, new Float32Array(i));
+        }
+      }
+      if (!w.faceImages.has(l.face_image)) {
+        const o = x.file(l.face_image);
+        if (o) {
+          const i = await o.async("blob"), c = await createImageBitmap(i);
+          w.faceImages.set(l.face_image, c);
+        }
+      }
+      if (!w.fullImages.has(l.full_image)) {
+        const o = x.file(l.full_image);
+        if (o) {
+          const i = await o.async("blob");
+          w.fullImages.set(l.full_image, i);
+        }
+      }
+    });
+    await Promise.all(n), x = null;
+    const a = [];
+    Z() && a.push("webgpu"), a.push("wasm");
+    const s = new URL(t.modelPath, self.location.href).toString(), u = new URL(s).pathname.split("/").pop() || "model.onnx", d = new URL(\`\${u}.data\`, s).toString();
+    _ = await $.create(s, {
+      executionProviders: a,
+      graphOptimizationLevel: "all",
+      enableMemPattern: !0,
+      enableCpuMemArena: !0,
+      externalData: [{ data: d, path: \`\${u}.data\` }]
+    }), self.postMessage({ type: "ready" });
+  } catch (e) {
+    T(e?.message || "Init Error");
+  }
+}, oe = async (t) => {
+  if (!_ || !R) {
+    T("Not initialized");
+    return;
+  }
+  P = !1;
+  const e = t.audioDimensions[0];
+  self.postMessage({
+    type: "start",
+    payload: { totalFrames: e, frameInterval: j }
+  });
+  let n = 0, a = 1;
+  const s = {}, u = _.inputNames[0], d = _.inputNames[1], l = _.outputNames[0];
+  v = "", p && (p.close(), p = null);
+  for (let o = 0; o < e && !P; o++) {
+    let i = null, c = null, r = null, m = null;
+    try {
+      const f = R.images[n];
+      i = K(f.tensor_file), f.full_image !== v && (p && p.close(), p = await Y(f.full_image), v = f.full_image);
+      const h = ae(
+        t.audioFeatures,
+        t.audioDimensions,
+        o
+      );
+      c = new W("float32", h, [
+        1,
+        F,
+        G,
+        V
+      ]), s[u] = i, s[d] = c, r = await _.run(s);
+      const N = r[l];
+      if (!N)
+        throw new Error("Missing inference output tensor");
+      m = N;
+      const B = await te(N, f, p);
+      self.postMessage(
+        { type: "frame", payload: { frame: B, index: o } },
+        [B]
+      );
+      const U = Q(
+        n,
+        R.images.length,
+        a
+      );
+      n = U.nextIndex, a = U.nextDirection;
+    } catch (f) {
+      T(f?.message || "Run Error");
+      break;
+    } finally {
+      i && i.dispose(), c && c.dispose(), m && m.dispose(), i = null, c = null, m = null, r = null;
     }
-  );
+  }
+  p && (p.close(), p = null), v = "", self.postMessage({ type: "done" });
+};
+self.onmessage = (t) => {
+  const e = t.data;
+  e.type === "init" ? se(e) : e.type === "run" ? oe(e) : e.type === "stop" && (P = !0);
+};
+`, m = typeof self < "u" && self.Blob && new Blob(["URL.revokeObjectURL(import.meta.url);", d], { type: "text/javascript;charset=utf-8" });
+function w(a) {
+  let n;
+  try {
+    if (n = m && (self.URL || self.webkitURL).createObjectURL(m), !n) throw "";
+    const e = new Worker(n, {
+      type: "module",
+      name: a?.name
+    });
+    return e.addEventListener("error", () => {
+      (self.URL || self.webkitURL).revokeObjectURL(n);
+    }), e;
+  } catch {
+    return new Worker(
+      "data:text/javascript;charset=utf-8," + encodeURIComponent(d),
+      {
+        type: "module",
+        name: a?.name
+      }
+    );
+  }
 }
-class v {
+class O {
   worker;
   handlers = {};
-  constructor(e = {}) {
-    this.handlers = e, this.worker = new E(), this.worker.onmessage = this.handleMessage.bind(this);
+  constructor(n = {}) {
+    this.handlers = n, this.worker = new w(), this.worker.onmessage = this.handleMessage.bind(this);
   }
-  handleMessage(e) {
-    const { type: s, payload: r } = e.data;
-    switch (s) {
+  handleMessage(n) {
+    const { type: e, payload: t } = n.data;
+    switch (e) {
       case "ready":
         this.handlers.onReady?.();
         break;
       case "start":
-        this.handlers.onStart?.(r.totalFrames);
+        this.handlers.onStart?.(t.totalFrames);
         break;
       case "frame":
-        this.handlers.onFrame?.(r.frame, r.index);
+        this.handlers.onFrame?.(t.frame, t.index);
         break;
       case "done":
         this.handlers.onDone?.();
         break;
       case "error":
-        this.handlers.onError?.(r);
+        this.handlers.onError?.(t);
         break;
     }
   }
-  init(e) {
+  init(n) {
     this.worker.postMessage(
       {
         type: "init",
-        modelPath: e.modelPath,
-        dataset: e.dataset,
-        zipBuffer: e.zipBuffer,
-        blendingMask: e.blendingMask,
-        wasmPaths: e.wasmPaths
+        modelPath: n.modelPath,
+        dataset: n.dataset,
+        zipBuffer: n.zipBuffer,
+        blendingMask: n.blendingMask,
+        wasmPaths: n.wasmPaths
       },
-      [e.zipBuffer, e.blendingMask]
+      [n.zipBuffer, n.blendingMask]
     );
   }
-  run(e) {
+  run(n) {
     this.worker.postMessage(
       {
         type: "run",
-        audioFeatures: e.audioFeatures,
-        audioDimensions: e.audioDimensions
+        audioFeatures: n.audioFeatures,
+        audioDimensions: n.audioDimensions
       },
-      [e.audioFeatures.buffer]
+      [n.audioFeatures.buffer]
     );
   }
   stop() {
@@ -116,21 +465,21 @@ class v {
   }
 }
 export {
-  g as DEFAULT_LOW_FREQ,
-  _ as ENERGY_FLOOR,
-  L as FRAME_INTERVAL_MS,
-  m as FRAME_LENGTH_MS,
-  u as FRAME_SHIFT_MS,
-  O as FeatureExtractor,
-  v as InferenceEngine,
-  k as NUM_MEL_BINS,
-  f as NUM_SEQUENCE_FRAMES,
-  p as PREEMPH_COEFF,
-  M as REMOVE_DC_OFFSET,
-  R as ROUND_TO_POWER_OF_TWO,
-  w as SAMPLE_RATE,
-  F as TARGET_FPS,
-  P as calculatePingPongState,
-  y as getBorder,
-  b as loadDataset
+  k as DEFAULT_LOW_FREQ,
+  R as ENERGY_FLOOR,
+  x as FRAME_INTERVAL_MS,
+  y as FRAME_LENGTH_MS,
+  _ as FRAME_SHIFT_MS,
+  A as FeatureExtractor,
+  O as InferenceEngine,
+  M as NUM_MEL_BINS,
+  E as NUM_SEQUENCE_FRAMES,
+  b as PREEMPH_COEFF,
+  I as REMOVE_DC_OFFSET,
+  v as ROUND_TO_POWER_OF_TWO,
+  g as SAMPLE_RATE,
+  L as TARGET_FPS,
+  C as calculatePingPongState,
+  F as getBorder,
+  U as loadDataset
 };

package/package.json

@@ -1,42 +1,41 @@
-{
-  "name": "sophontalk-services",
-  "version": "0.0.2",
-  "type": "module",
-  "scripts": {
-    "dev": "vite",
-    "build": "vue-tsc -b && vite build",
-    "build:lib": "vite build --config vite.lib.config.ts",
-    "preview": "vite preview",
-    "publish": "npm publish  --registry https://registry.npmjs.org --access public"
-  },
-  "main": "./dist-lib/sophontalk-services.cjs",
-  "module": "./dist-lib/sophontalk-services.js",
-  "types": "./dist-lib/types/src/services/index.d.ts",
-  "files": [
-    "dist-lib"
-  ],
-  "exports": {
-    ".": {
-      "types": "./dist-lib/types/src/services/index.d.ts",
-      "import": "./dist-lib/sophontalk-services.js",
-      "require": "./dist-lib/sophontalk-services.cjs"
-    }
-  },
-  "dependencies": {
-    "@alexanderolsen/libsamplerate-js": "^2.1.2",
-    "fft.js": "^4.0.4",
-    "jszip": "^3.10.1",
-    "onnxruntime-web": "^1.22.0",
-    "vue": "^3.5.17"
-  },
-  "devDependencies": {
-    "@rollup/plugin-replace": "^6.0.3",
-    "@vitejs/plugin-vue": "^6.0.0",
-    "@vue/tsconfig": "^0.7.0",
-    "typescript": "~5.8.3",
-    "vite": "^7.0.2",
-    "vite-plugin-dts": "^4.5.4",
-    "vite-plugin-static-copy": "^3.1.0",
-    "vue-tsc": "^3.0.1"
-  }
-}
+{
+  "name": "sophontalk-services",
+  "version": "0.0.4",
+  "type": "module",
+  "scripts": {
+    "dev": "vite",
+    "build": "vue-tsc -b && vite build",
+    "build:lib": "vite build --config vite.lib.config.ts",
+    "preview": "vite preview"
+  },
+  "main": "./dist-lib/sophontalk-services.cjs",
+  "module": "./dist-lib/sophontalk-services.js",
+  "types": "./dist-lib/types/src/services/index.d.ts",
+  "files": [
+    "dist-lib"
+  ],
+  "exports": {
+    ".": {
+      "types": "./dist-lib/types/src/services/index.d.ts",
+      "import": "./dist-lib/sophontalk-services.js",
+      "require": "./dist-lib/sophontalk-services.cjs"
+    }
+  },
+  "dependencies": {
+    "@alexanderolsen/libsamplerate-js": "^2.1.2",
+    "fft.js": "^4.0.4",
+    "jszip": "^3.10.1",
+    "onnxruntime-web": "^1.22.0",
+    "vue": "^3.5.17"
+  },
+  "devDependencies": {
+    "@rollup/plugin-replace": "^6.0.3",
+    "@vitejs/plugin-vue": "^6.0.0",
+    "@vue/tsconfig": "^0.7.0",
+    "typescript": "~5.8.3",
+    "vite": "^7.0.2",
+    "vite-plugin-dts": "^4.5.4",
+    "vite-plugin-static-copy": "^3.1.0",
+    "vue-tsc": "^3.0.1"
+  }
+}

package/dist-lib/assets/feature.worker-Dx6moind.js

@@ -1,145 +0,0 @@
-import x from "fft.js";
-import { create as O, ConverterType as H } from "@alexanderolsen/libsamplerate-js";
-const y = 16e3, U = 25, D = 10, u = 80, S = 4, W = 0.97, L = 1e-10, b = 20, h = Math.round(U * y / 1e3), C = Math.round(D * y / 1e3), k = (e) => {
-  if (e <= 0) return 1;
-  let t = 1;
-  for (; t < e; ) t *= 2;
-  return t;
-}, P = k(h), N = new x(P), v = new Float32Array(h), G = () => {
-  if (h <= 1) return;
-  const e = 2 * Math.PI / (h - 1);
-  for (let t = 0; t < h; t++)
-    v[t] = 0.5 - 0.5 * Math.cos(t * e);
-};
-G();
-function Q(e, t, o, a = 20, s) {
-  s = s || o / 2;
-  const i = (n) => 700 * (Math.exp(n / 1127) - 1), r = (n) => 1127 * Math.log(1 + n / 700), l = r(a), A = r(s), w = new Float32Array(e + 2);
-  for (let n = 0; n < e + 2; n++)
-    w[n] = l + (A - l) * n / (e + 1);
-  const _ = w.map(i).map((n) => n * t / o), f = Math.floor(t / 2) + 1, m = [];
-  for (let n = 1; n <= e; n++) {
-    const c = _[n - 1], F = _[n], M = _[n + 1], I = Math.ceil(c), T = Math.floor(M), E = Math.max(0, I), p = Math.min(f - 1, T), B = p - E + 1;
-    if (B <= 0) {
-      m.push({ startBin: 0, values: new Float32Array(0) });
-      continue;
-    }
-    const R = new Float32Array(B);
-    for (let d = E; d <= p; d++) {
-      let g = 0;
-      d <= F ? g = (d - c) / (F - c) : g = (M - d) / (M - F), R[d - E] = g;
-    }
-    m.push({ startBin: E, values: R });
-  }
-  return m;
-}
-const q = Q(
-  u,
-  P,
-  y,
-  b
-);
-function K(e) {
-  for (let t = 0; t < e.length; t++)
-    e[t] *= v[t];
-}
-function Y(e, t) {
-  for (let o = e.length - 1; o > 0; o--)
-    e[o] -= t * e[o - 1];
-  e[0] -= t * e[0];
-}
-function V(e) {
-  const t = e.length;
-  if (t === 0) return;
-  let o = 0;
-  for (let s = 0; s < t; s++) o += e[s];
-  const a = o / t;
-  for (let s = 0; s < t; s++) e[s] -= a;
-}
-function Z(e, t) {
-  const o = e.length;
-  for (let a = 0; a < o; a++) {
-    const s = t[a * 2], i = t[a * 2 + 1];
-    e[a] = s * s + i * i;
-  }
-}
-function j(e) {
-  const t = e.length;
-  if (t < 2) return;
-  let o = 0;
-  for (let r = 0; r < t; r++) o += e[r];
-  const a = o / t;
-  let s = 0;
-  for (let r = 0; r < t; r++) {
-    const l = e[r] - a;
-    s += l * l;
-  }
-  const i = Math.sqrt(s / (t - 1));
-  if (i > 1e-8) {
-    const r = 1 / i;
-    for (let l = 0; l < t; l++)
-      e[l] = (e[l] - a) * r;
-  } else
-    for (let r = 0; r < t; r++)
-      e[r] -= a;
-}
-async function z(e) {
-  let t = e.leftChannel;
-  if (e.rightChannel) {
-    const f = e.rightChannel, m = new Float32Array(t.length);
-    for (let n = 0; n < t.length; n++)
-      m[n] = (t[n] + f[n]) * 0.5;
-    t = m;
-  }
-  if (e.sampleRate !== y) {
-    const f = await O(1, e.sampleRate, y, {
-      converterType: H.SRC_SINC_MEDIUM_QUALITY
-    });
-    t = f.simple(t), f.destroy();
-  }
-  const o = Math.floor((t.length - h) / C) + 1;
-  if (o <= 0)
-    throw new Error("特征提取失败：音频时长过短。");
-  const a = Math.ceil(o / S) * S, s = a * u, i = new Float32Array(s), r = new Float32Array(h), l = N.createComplexArray(), A = N.createComplexArray(), w = new Float32Array(Math.floor(P / 2) + 1);
-  for (let f = 0; f < o; f++) {
-    const m = f * C;
-    r.set(t.subarray(m, m + h)), V(r), Y(r, W), K(r), l.fill(0);
-    for (let c = 0; c < h; c++)
-      l[c * 2] = r[c];
-    N.transform(A, l), Z(w, A);
-    const n = f * u;
-    for (let c = 0; c < u; c++) {
-      const F = q[c];
-      let M = 0;
-      const I = F.values, T = F.startBin, E = I.length;
-      for (let p = 0; p < E; p++)
-        M += w[T + p] * I[p];
-      M = M < L ? L : M, i[n + c] = Math.log(M);
-    }
-  }
-  if (a > o) {
-    const f = (o - 1) * u, m = f + u, n = i.subarray(f, m);
-    for (let c = o; c < a; c++)
-      i.set(n, c * u);
-  }
-  j(i);
-  const _ = a / S;
-  return {
-    features: i,
-    dimensions: [_, S, u]
-  };
-}
-self.addEventListener(
-  "message",
-  async (e) => {
-    try {
-      const t = e.data, o = await z(t);
-      self.postMessage(
-        { status: "success", payload: o },
-        { transfer: [o.features.buffer] }
-      );
-    } catch (t) {
-      self.postMessage({ status: "error", error: t.message });
-    }
-  }
-);

package/dist-lib/assets/inference.worker-cwIlIomt.js

@@ -1,175 +0,0 @@
-import { env as B, InferenceSession as V, Tensor as U } from "onnxruntime-web";
-import $ from "jszip";
-const G = 80, L = 4, H = 25, j = 1e3 / H, J = (t) => {
-  if (t === 128) return 4;
-  if (t === 96) return 3;
-  throw new Error(`Unsupported crop size: ${t}. Only nano (96) and tiny (128) are supported.`);
-}, Q = (t, e, a) => {
-  if (e <= 1) return { nextIndex: 0, nextDirection: 1 };
-  let s = t + a, o = a;
-  return s >= e ? (s = e - 2, o = -1) : s < 0 && (s = 1, o = 1), { nextIndex: s, nextDirection: o };
-};
-B.wasm.numThreads = 1;
-const Z = () => typeof navigator < "u" && !!navigator.gpu, F = 32;
-let _ = null, R = null, M = null, q = null;
-const w = {
-  tensors: /* @__PURE__ */ new Map(),
-  faceImages: /* @__PURE__ */ new Map(),
-  fullImages: /* @__PURE__ */ new Map()
-};
-let P = 3, b = 96, k = 0, W = 0, C = null, N = null, y = null, E = null, g = null, I = null, S = null, x = null, T = !1, v = "", p = null;
-const D = (t) => {
-  self.postMessage({ type: "error", payload: t });
-}, K = (t) => {
-  const e = w.tensors.get(t);
-  if (!e) throw new Error(`Cache miss: ${t}`);
-  const a = b - 2 * P;
-  return new U("float32", e, [1, 6, a, a]);
-}, X = (t) => {
-  const e = w.faceImages.get(t);
-  if (!e) throw new Error(`Cache miss: ${t}`);
-  return e;
-}, Y = async (t) => {
-  const e = w.fullImages.get(t);
-  if (!e) throw new Error(`Cache miss: ${t}`);
-  return await createImageBitmap(e);
-}, ee = (t) => {
-  const [e, a, s, o] = t.dims, c = t.data;
-  (!y || y.width !== o || y.height !== s) && (C.width = o, C.height = s, y = N.createImageData(o, s));
-  const n = s * o, l = new Uint32Array(y.data.buffer), i = n, d = n * 2, u = a === 3;
-  for (let r = 0; r < n; r++) {
-    let m = (u ? c[r + d] : c[r]) * 255, f = (u ? c[r + i] : c[r]) * 255, h = c[r] * 255;
-    m = m < 0 ? 0 : m > 255 ? 255 : m, f = f < 0 ? 0 : f > 255 ? 255 : f, h = h < 0 ? 0 : h > 255 ? 255 : h, l[r] = 255 << 24 | (h & 255) << 16 | (f & 255) << 8 | m & 255;
-  }
-  N.putImageData(y, 0, 0);
-}, te = async (t, e, a) => {
-  ee(t);
-  const s = X(e.face_image);
-  g.globalCompositeOperation = "source-over", g.drawImage(s, 0, 0), g.drawImage(C, P, P), g.globalCompositeOperation = "destination-in", g.drawImage(q, 0, 0, b, b), g.globalCompositeOperation = "destination-over", g.drawImage(s, 0, 0), g.globalCompositeOperation = "source-over", S.drawImage(a, 0, 0);
-  const o = e.crop_info.ymax - e.crop_info.ymin;
-  return S.drawImage(
-    E,
-    e.crop_info.xmin,
-    e.crop_info.ymin,
-    e.crop_info.width,
-    o
-  ), await createImageBitmap(I);
-}, ae = (t, e, a) => {
-  const [s, o, c] = e, n = o * c, l = F * n;
-  (!x || x.length !== l) && (x = new Float32Array(l));
-  for (let i = 0; i < F; i++) {
-    const u = Math.min(
-      Math.max(a - F / 2 + i, 0),
-      s - 1
-    ) * n, r = i * n;
-    x.set(
-      t.subarray(u, u + n),
-      r
-    );
-  }
-  return x;
-}, ne = (t, e, a) => {
-  C || (C = new OffscreenCanvas(1, 1), N = C.getContext("2d", {
-    willReadFrequently: !0
-  })), (!E || E.width !== a) && (E = new OffscreenCanvas(a, a), g = E.getContext("2d", {
-    willReadFrequently: !0
-  })), (!I || I.width !== t || I.height !== e) && (I = new OffscreenCanvas(t, e), S = I.getContext("2d", {
-    willReadFrequently: !0
-  }));
-}, se = async (t) => {
-  try {
-    if (t.wasmPaths ? B.wasm.wasmPaths = t.wasmPaths : B.wasm.wasmPaths = "/", !t.dataset) throw new Error("Missing dataset");
-    const e = t.dataset;
-    R = e, b = e.dataset_info.config.crop_size, P = J(b), k = e.dataset_info.source_image_dimensions?.width || 0, W = e.dataset_info.source_image_dimensions?.height || 0, ne(k, W, b), M = await $.loadAsync(t.zipBuffer), t.zipBuffer = null, q = t.blendingMask;
-    const a = e.images.map(async (n) => {
-      if (!w.tensors.has(n.tensor_file)) {
-        const l = M.file(n.tensor_file);
-        if (l) {
-          const i = await l.async("arraybuffer");
-          w.tensors.set(n.tensor_file, new Float32Array(i));
-        }
-      }
-      if (!w.faceImages.has(n.face_image)) {
-        const l = M.file(n.face_image);
-        if (l) {
-          const i = await l.async("blob"), d = await createImageBitmap(i);
-          w.faceImages.set(n.face_image, d);
-        }
-      }
-      if (!w.fullImages.has(n.full_image)) {
-        const l = M.file(n.full_image);
-        if (l) {
-          const i = await l.async("blob");
-          w.fullImages.set(n.full_image, i);
-        }
-      }
-    });
-    await Promise.all(a), M = null;
-    const o = await (await fetch(t.modelPath)).arrayBuffer(), c = [];
-    Z() && c.push("webgpu"), c.push("wasm"), _ = await V.create(o, {
-      executionProviders: c,
-      graphOptimizationLevel: "all",
-      enableMemPattern: !0,
-      enableCpuMemArena: !0
-    }), self.postMessage({ type: "ready" });
-  } catch (e) {
-    D(e?.message || "Init Error");
-  }
-}, oe = async (t) => {
-  if (!_ || !R) {
-    D("Not initialized");
-    return;
-  }
-  T = !1;
-  const e = t.audioDimensions[0];
-  self.postMessage({
-    type: "start",
-    payload: { totalFrames: e, frameInterval: j }
-  });
-  let a = 0, s = 1;
-  const o = {}, c = _.inputNames[0], n = _.inputNames[1], l = _.outputNames[0];
-  v = "", p && (p.close(), p = null);
-  for (let i = 0; i < e && !T; i++) {
-    let d = null, u = null, r = null, m = null;
-    try {
-      const f = R.images[a];
-      d = K(f.tensor_file), f.full_image !== v && (p && p.close(), p = await Y(f.full_image), v = f.full_image);
-      const h = ae(
-        t.audioFeatures,
-        t.audioDimensions,
-        i
-      );
-      u = new U("float32", h, [
-        1,
-        F,
-        L,
-        G
-      ]), o[c] = d, o[n] = u, r = await _.run(o);
-      const A = r[l];
-      if (!A)
-        throw new Error("Missing inference output tensor");
-      m = A;
-      const O = await te(A, f, p);
-      self.postMessage(
-        { type: "frame", payload: { frame: O, index: i } },
-        [O]
-      );
-      const z = Q(
-        a,
-        R.images.length,
-        s
-      );
-      a = z.nextIndex, s = z.nextDirection;
-    } catch (f) {
-      D(f?.message || "Run Error");
-      break;
-    } finally {
-      d && d.dispose(), u && u.dispose(), m && m.dispose(), d = null, u = null, m = null, r = null;
-    }
-  }
-  p && (p.close(), p = null), v = "", self.postMessage({ type: "done" });
-};
-self.onmessage = (t) => {
-  const e = t.data;
-  e.type === "init" ? se(e) : e.type === "run" ? oe(e) : e.type === "stop" && (T = !0);
-};