sophontalk-services 0.0.4 → 0.0.5

package/dist-lib/sophontalk-services.js

@@ -1,460 +1,108 @@
-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 };
+import { env as l } from "onnxruntime-web";
+const m = 16e3, k = 25, u = 10, f = 80, p = 4, g = 0.97, _ = 1e-10, M = 20, R = !0, F = !0, w = 25, L = 1e3 / w, y = (s) => {
+  if (s === 128) return 4;
+  if (s === 96) return 3;
+  throw new Error(`Unsupported crop size: ${s}. Only nano (96) and tiny (128) are supported.`);
+}, P = (s, e, t) => {
+  if (e <= 1) return { nextIndex: 0, nextDirection: 1 };
+  let r = s + t, a = t;
+  return r >= e ? (r = e - 2, a = -1) : r < 0 && (r = 1, a = 1), { nextIndex: r, nextDirection: a };
 };
-async function U(a = {}) {
+async function b(s = {}) {
   const {
-    jsonUrl: n = "/complete_dataset.json",
-    zipUrl: e = "/processed_images.zip"
-  } = a, [t, r] = await Promise.all([
-    fetch(n),
-    fetch(e)
+    jsonUrl: e = "/complete_dataset.json",
+    zipUrl: t = "/processed_images.zip"
+  } = s, [r, a] = await Promise.all([
+    fetch(e),
+    fetch(t)
   ]);
-  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();
+  if (!a.ok) throw new Error(`无法加载 ${t}`);
+  const o = await r.json(), c = await a.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 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);
+class v {
+  worker = new Worker(
+    new URL(
+      /* @vite-ignore */
+      "./workers/feature.worker.js",
+      import.meta.url
+    ),
+    { type: "module" }
+  );
+  async process(e) {
+    const t = e.getChannelData(0), r = e.numberOfChannels > 1 ? e.getChannelData(1) : void 0, a = {
+      leftChannel: t,
+      rightChannel: r,
+      sampleRate: e.sampleRate
+    }, o = [t.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", h), c(n.data.payload)) : n.data?.status === "error" && (this.worker.removeEventListener("message", i), this.worker.removeEventListener("error", h), d(new Error(n.data.error || "特征提取失败")));
+      }, h = (n) => {
+        this.worker.removeEventListener("message", i), this.worker.removeEventListener("error", h), d(n);
       };
-      this.worker.addEventListener("message", l), this.worker.addEventListener("error", i), this.worker.postMessage(r, o);
+      this.worker.addEventListener("message", i), this.worker.addEventListener("error", h), this.worker.postMessage(a, o);
     });
   }
   dispose() {
     this.worker.terminate();
   }
 }
-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 O {
   worker;
   handlers = {};
-  constructor(n = {}) {
-    this.handlers = n, this.worker = new w(), this.worker.onmessage = this.handleMessage.bind(this);
-  }
-  handleMessage(n) {
-    const { type: e, payload: t } = n.data;
-    switch (e) {
+  constructor(e = {}) {
+    this.handlers = e, this.worker = new Worker(
+      new URL(
+        /* @vite-ignore */
+        "./workers/inference.worker.js",
+        import.meta.url
+      ),
+      { type: "module" }
+    ), this.worker.onmessage = this.handleMessage.bind(this);
+  }
+  handleMessage(e) {
+    const { type: t, payload: r } = e.data;
+    switch (t) {
       case "ready":
         this.handlers.onReady?.();
         break;
       case "start":
-        this.handlers.onStart?.(t.totalFrames);
+        this.handlers.onStart?.(r.totalFrames);
         break;
       case "frame":
-        this.handlers.onFrame?.(t.frame, t.index);
+        this.handlers.onFrame?.(r.frame, r.index);
         break;
       case "done":
         this.handlers.onDone?.();
         break;
       case "error":
-        this.handlers.onError?.(t);
+        this.handlers.onError?.(r);
         break;
     }
   }
-  init(n) {
+  init(e) {
     this.worker.postMessage(
       {
         type: "init",
-        modelPath: n.modelPath,
-        dataset: n.dataset,
-        zipBuffer: n.zipBuffer,
-        blendingMask: n.blendingMask,
-        wasmPaths: n.wasmPaths
+        modelPath: e.modelPath,
+        dataset: e.dataset,
+        zipBuffer: e.zipBuffer,
+        blendingMask: e.blendingMask,
+        wasmPaths: e.wasmPaths || l.wasm.wasmPaths
       },
-      [n.zipBuffer, n.blendingMask]
+      [e.zipBuffer, e.blendingMask]
     );
   }
-  run(n) {
+  run(e) {
     this.worker.postMessage(
       {
         type: "run",
-        audioFeatures: n.audioFeatures,
-        audioDimensions: n.audioDimensions
+        audioFeatures: e.audioFeatures,
+        audioDimensions: e.audioDimensions
       },
-      [n.audioFeatures.buffer]
+      [e.audioFeatures.buffer]
     );
   }
   stop() {
@@ -465,21 +113,21 @@ class O {
   }
 }
 export {
-  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,
+  M as DEFAULT_LOW_FREQ,
+  _ as ENERGY_FLOOR,
+  L as FRAME_INTERVAL_MS,
+  k as FRAME_LENGTH_MS,
+  u as FRAME_SHIFT_MS,
+  v 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
+  f as NUM_MEL_BINS,
+  p as NUM_SEQUENCE_FRAMES,
+  g as PREEMPH_COEFF,
+  R as REMOVE_DC_OFFSET,
+  F as ROUND_TO_POWER_OF_TWO,
+  m as SAMPLE_RATE,
+  w as TARGET_FPS,
+  P as calculatePingPongState,
+  y as getBorder,
+  b as loadDataset
 };