@omote/core 0.2.3 → 0.3.1

package/dist/index.mjs

@@ -13,10 +13,10 @@ import {
   setLoggingEnabled
 } from "./chunk-ESU52TDS.mjs";
 import {
-  env,
-  pipeline3
-} from "./chunk-RI6UQ7WF.mjs";
-import "./chunk-NSSMTXJJ.mjs";
+  __webpack_exports__env,
+  __webpack_exports__pipeline
+} from "./chunk-T465MTDX.mjs";
+import "./chunk-6W7G6WE7.mjs";
 
 // src/audio/MicrophoneCapture.ts
 var MicrophoneCapture = class {
@@ -263,7 +263,7 @@ var AudioScheduler = class {
     const ctx = await this.ensureContext();
     const channels = this.options.channels ?? 1;
     if (!this.isPlaying) {
-      this.nextPlayTime = ctx.currentTime + 0.05;
+      this.nextPlayTime = ctx.currentTime + (this.options.initialDelayS ?? 0.05);
       this.isPlaying = true;
     }
     const audioBuffer = ctx.createBuffer(channels, audioData.length, ctx.sampleRate);
@@ -446,8 +446,8 @@ var AudioChunkCoalescer = class {
 var LAMPipeline = class {
   constructor(options = {}) {
     this.options = options;
-    this.REQUIRED_SAMPLES = 16e3;
-    // 1.0s at 16kHz (LAM requirement)
+    this.DEFAULT_CHUNK_SAMPLES = 16e3;
+    // 1.0s at 16kHz (Wav2Vec2 requirement)
     this.FRAME_RATE = 30;
     // LAM outputs 30fps
     this.buffer = new Float32Array(0);
@@ -477,19 +477,20 @@ var LAMPipeline = class {
     newBuffer.set(this.buffer, 0);
     newBuffer.set(samples, this.buffer.length);
     this.buffer = newBuffer;
-    while (this.buffer.length >= this.REQUIRED_SAMPLES) {
-      await this.processBuffer(lam);
+    const chunkSize = lam.chunkSamples ?? this.DEFAULT_CHUNK_SAMPLES;
+    while (this.buffer.length >= chunkSize) {
+      await this.processBuffer(lam, chunkSize);
     }
   }
   /**
    * Process accumulated buffer through LAM inference
    */
-  async processBuffer(lam) {
+  async processBuffer(lam, chunkSize) {
     try {
-      const toProcess = this.buffer.slice(0, this.REQUIRED_SAMPLES);
+      const toProcess = this.buffer.slice(0, chunkSize);
       const processedStartTime = this.bufferStartTime;
-      this.buffer = this.buffer.slice(this.REQUIRED_SAMPLES);
-      const processedDuration = this.REQUIRED_SAMPLES / (this.options.sampleRate ?? 16e3);
+      this.buffer = this.buffer.slice(chunkSize);
+      const processedDuration = chunkSize / (this.options.sampleRate ?? 16e3);
       this.bufferStartTime = processedStartTime + processedDuration;
       const result = await lam.infer(toProcess);
       const frameDuration = 1 / this.FRAME_RATE;
@@ -508,35 +509,22 @@ var LAMPipeline = class {
   /**
    * Get the frame that should be displayed at the current time
    *
-   * Automatically removes frames that have already been displayed.
-   * This prevents memory leaks from accumulating old frames.
+   * Timestamp-synced playback for all backends. Audio playback is delayed
+   * for slow backends (WASM gets 1s head start via AudioScheduler) so
+   * frames are ready by the time their corresponding audio plays.
    *
-   * Discard Window (prevents premature frame discarding):
-   * - WebGPU: 0.5s (LAM inference 20-100ms + RAF jitter + React stalls)
-   * - WASM: 1.0s (LAM inference 50-500ms + higher variability)
-   *
-   * Last-Frame-Hold: Returns last valid frame instead of null to prevent
-   * avatar freezing when between frames (RAF at 60fps vs LAM at 30fps).
+   * Discard window is generous for WASM to handle inference jitter.
+   * Late frames play at RAF rate (~60fps) until caught up, then settle
+   * to natural 30fps pacing via timestamp gating.
    *
    * @param currentTime - Current AudioContext time
    * @param lam - LAM inference engine (optional, for backend detection)
    * @returns Current frame, or last frame as fallback, or null if no frames yet
    */
   getFrameForTime(currentTime, lam) {
-    const discardWindow = lam?.backend === "wasm" ? 1 : 0.5;
-    let discardedCount = 0;
+    const discardWindow = lam?.backend === "wasm" ? 10 : 0.5;
     while (this.frameQueue.length > 0 && this.frameQueue[0].timestamp < currentTime - discardWindow) {
-      const discarded = this.frameQueue.shift();
-      discardedCount++;
-      if (discardedCount === 1) {
-        const ageMs = ((currentTime - discarded.timestamp) * 1e3).toFixed(0);
-        console.warn("[LAM] Frame(s) discarded as too old", {
-          ageMs,
-          discardWindowMs: discardWindow * 1e3,
-          queueLength: this.frameQueue.length,
-          backend: lam?.backend ?? "unknown"
-        });
-      }
+      this.frameQueue.shift();
     }
     if (this.frameQueue.length > 0 && this.frameQueue[0].timestamp <= currentTime) {
       const { frame } = this.frameQueue.shift();
@@ -555,7 +543,7 @@ var LAMPipeline = class {
    * Get current buffer fill level (0-1)
    */
   get fillLevel() {
-    return Math.min(1, this.buffer.length / this.REQUIRED_SAMPLES);
+    return Math.min(1, this.buffer.length / this.DEFAULT_CHUNK_SAMPLES);
   }
   /**
    * Get number of frames queued
@@ -572,7 +560,7 @@ var LAMPipeline = class {
   /**
    * Flush remaining buffered audio
    *
-   * Processes any remaining audio in the buffer, even if less than REQUIRED_SAMPLES.
+   * Processes any remaining audio in the buffer, even if less than the chunk size.
    * This ensures the final audio chunk generates blendshape frames.
    *
    * Should be called when audio stream ends to prevent losing the last 0-1 seconds.
@@ -583,12 +571,17 @@ var LAMPipeline = class {
     if (this.buffer.length === 0) {
       return;
     }
-    const padded = new Float32Array(this.REQUIRED_SAMPLES);
-    padded.set(this.buffer, 0);
     const processedStartTime = this.bufferStartTime;
+    const sampleRate = this.options.sampleRate ?? 16e3;
+    const minSize = lam.chunkSamples ?? this.DEFAULT_CHUNK_SAMPLES;
+    const audioToInfer = this.buffer.length >= minSize ? this.buffer : (() => {
+      const padded = new Float32Array(minSize);
+      padded.set(this.buffer, 0);
+      return padded;
+    })();
     try {
-      const result = await lam.infer(padded);
-      const actualDuration = this.buffer.length / (this.options.sampleRate ?? 16e3);
+      const result = await lam.infer(audioToInfer);
+      const actualDuration = this.buffer.length / sampleRate;
       const frameDuration = 1 / this.FRAME_RATE;
       const actualFrameCount = Math.ceil(actualDuration * this.FRAME_RATE);
       for (let i = 0; i < Math.min(actualFrameCount, result.blendshapes.length); i++) {
@@ -647,7 +640,13 @@ var SyncedAudioPipeline = class extends EventEmitter {
     this.monitorInterval = null;
     this.frameAnimationId = null;
     const sampleRate = options.sampleRate ?? 16e3;
-    this.scheduler = new AudioScheduler({ sampleRate });
+    if (!options.lam.isLoaded) {
+      throw new Error(
+        "LipSyncBackend must be loaded before constructing SyncedAudioPipeline. Call lam.load() first so backend type is known for timing configuration."
+      );
+    }
+    const initialDelayS = options.lam.backend === "wasm" ? 1 : 0.05;
+    this.scheduler = new AudioScheduler({ sampleRate, initialDelayS });
     this.coalescer = new AudioChunkCoalescer({
       sampleRate,
       targetDurationMs: options.chunkTargetMs ?? 200
@@ -2138,12 +2137,9 @@ function applyIOSWasmMemoryPatch() {
   iosWasmPatched = true;
   const OrigMemory = WebAssembly.Memory;
   const MAX_IOS_PAGES = 16384;
-  logger.info("Applying iOS WASM memory patch (shared\u2192false, max\u21921GB)");
+  logger.info("Applying iOS WASM memory patch (max capped to 1GB, shared preserved)");
   WebAssembly.Memory = function IOSPatchedMemory(descriptor) {
     const patched = { ...descriptor };
-    if (patched.shared) {
-      patched.shared = false;
-    }
     if (patched.maximum !== void 0 && patched.maximum > MAX_IOS_PAGES) {
       patched.maximum = MAX_IOS_PAGES;
     }
@@ -2760,7 +2756,7 @@ var WhisperInference = class _WhisperInference {
    * Check if WebGPU is available in this browser
    */
   static async isWebGPUAvailable() {
-    return "gpu" in navigator;
+    return isWebGPUAvailable();
   }
   /**
    * Load the Whisper model pipeline
@@ -2801,19 +2797,19 @@ var WhisperInference = class _WhisperInference {
       const hasWebGPU = await _WhisperInference.isWebGPUAvailable();
       const device = this.config.device === "auto" ? hasWebGPU ? "webgpu" : "wasm" : this.config.device;
       logger4.info("Creating pipeline", { device, hasWebGPU });
-      env.allowLocalModels = false;
-      env.allowRemoteModels = true;
-      env.useBrowserCache = false;
-      env.useCustomCache = false;
-      env.useWasmCache = false;
-      if (env.backends.onnx.wasm) {
-        env.backends.onnx.wasm.proxy = false;
-        env.backends.onnx.wasm.numThreads = 1;
+      __webpack_exports__env.allowLocalModels = false;
+      __webpack_exports__env.allowRemoteModels = true;
+      __webpack_exports__env.useBrowserCache = false;
+      __webpack_exports__env.useCustomCache = false;
+      __webpack_exports__env.useWasmCache = false;
+      if (__webpack_exports__env.backends.onnx.wasm) {
+        __webpack_exports__env.backends.onnx.wasm.proxy = false;
+        __webpack_exports__env.backends.onnx.wasm.numThreads = 1;
       }
       logger4.info("Configured transformers.js env", {
-        allowLocalModels: env.allowLocalModels,
-        useBrowserCache: env.useBrowserCache,
-        useWasmCache: env.useWasmCache
+        allowLocalModels: __webpack_exports__env.allowLocalModels,
+        useBrowserCache: __webpack_exports__env.useBrowserCache,
+        useWasmCache: __webpack_exports__env.useWasmCache
       });
       const pipelineOptions = {
         dtype: this.config.dtype,
@@ -2830,7 +2826,7 @@ var WhisperInference = class _WhisperInference {
         };
         logger4.info("Forcing WebGPU execution providers");
       }
-      this.pipeline = await pipeline3(
+      this.pipeline = await __webpack_exports__pipeline(
         "automatic-speech-recognition",
         modelName,
         pipelineOptions
@@ -3061,6 +3057,12 @@ var Wav2ArkitCpuInference = class {
     this.isLoading = false;
     // Inference queue for handling concurrent calls
     this.inferenceQueue = Promise.resolve();
+    /**
+     * Preferred chunk size: 4000 samples (250ms at 16kHz).
+     * wav2arkit_cpu accepts variable-length input, so we use smaller chunks
+     * for lower latency on WASM (vs 16000 for Wav2Vec2's fixed requirement).
+     */
+    this.chunkSamples = 4e3;
     this.config = config;
   }
   get backend() {
@@ -3093,32 +3095,78 @@ var Wav2ArkitCpuInference = class {
       this.ort = ort;
       this._backend = backend;
       logger5.info("ONNX Runtime loaded", { backend: this._backend });
-      const cache = getModelCache();
       const modelUrl = this.config.modelUrl;
-      const isCached = await cache.has(modelUrl);
-      let modelBuffer;
-      if (isCached) {
-        logger5.debug("Loading model from cache", { modelUrl });
-        modelBuffer = await cache.get(modelUrl);
-        if (!modelBuffer) {
-          logger5.warn("Cache corruption detected, clearing and retrying", { modelUrl });
-          await cache.delete(modelUrl);
+      const sessionOptions = { ...getSessionOptions(this._backend) };
+      let isCached = false;
+      if (isIOS() && this.config.modelDataUrl) {
+        const dataFilename = this.config.modelDataUrl.split("/").pop();
+        sessionOptions.externalData = [{
+          path: dataFilename,
+          data: this.config.modelDataUrl
+        }];
+        logger5.info("iOS: URL-based session creation (ORT handles fetch internally)", {
+          modelUrl,
+          dataFile: dataFilename,
+          dataUrl: this.config.modelDataUrl
+        });
+        this.session = await this.ort.InferenceSession.create(modelUrl, sessionOptions);
+      } else {
+        const cache = getModelCache();
+        isCached = await cache.has(modelUrl);
+        let modelBuffer;
+        if (isCached) {
+          logger5.debug("Loading model from cache", { modelUrl });
+          modelBuffer = await cache.get(modelUrl);
+          if (!modelBuffer) {
+            logger5.warn("Cache corruption detected, clearing and retrying", { modelUrl });
+            await cache.delete(modelUrl);
+            modelBuffer = await fetchWithCache(modelUrl);
+          }
+        } else {
+          logger5.debug("Fetching and caching model", { modelUrl });
           modelBuffer = await fetchWithCache(modelUrl);
         }
-      } else {
-        logger5.debug("Fetching and caching model", { modelUrl });
-        modelBuffer = await fetchWithCache(modelUrl);
-      }
-      if (!modelBuffer) {
-        throw new Error(`Failed to load model: ${modelUrl}`);
+        if (!modelBuffer) {
+          throw new Error(`Failed to load model: ${modelUrl}`);
+        }
+        let externalDataBuffer;
+        if (this.config.modelDataUrl) {
+          const dataUrl = this.config.modelDataUrl;
+          const isDataCached = await cache.has(dataUrl);
+          if (isDataCached) {
+            logger5.debug("Loading external data from cache", { dataUrl });
+            externalDataBuffer = await cache.get(dataUrl);
+            if (!externalDataBuffer) {
+              logger5.warn("External data cache corruption, re-fetching", { dataUrl });
+              await cache.delete(dataUrl);
+              externalDataBuffer = await fetchWithCache(dataUrl);
+            }
+          } else {
+            logger5.info("Fetching external data (this may take a while on first load)", {
+              dataUrl
+            });
+            externalDataBuffer = await fetchWithCache(dataUrl);
+          }
+          logger5.debug("External data loaded", {
+            size: formatBytes(externalDataBuffer.byteLength)
+          });
+        }
+        logger5.debug("Creating ONNX session", {
+          size: formatBytes(modelBuffer.byteLength),
+          hasExternalData: !!externalDataBuffer,
+          externalDataSize: externalDataBuffer ? formatBytes(externalDataBuffer.byteLength) : void 0,
+          backend: this._backend
+        });
+        if (externalDataBuffer) {
+          const dataFilename = this.config.modelDataUrl.split("/").pop();
+          sessionOptions.externalData = [{
+            path: dataFilename,
+            data: new Uint8Array(externalDataBuffer)
+          }];
+        }
+        const modelData = new Uint8Array(modelBuffer);
+        this.session = await this.ort.InferenceSession.create(modelData, sessionOptions);
       }
-      logger5.debug("Creating ONNX session", {
-        size: formatBytes(modelBuffer.byteLength),
-        backend: this._backend
-      });
-      const sessionOptions = getSessionOptions(this._backend);
-      const modelData = new Uint8Array(modelBuffer);
-      this.session = await this.ort.InferenceSession.create(modelData, sessionOptions);
       const loadTimeMs = performance.now() - startTime;
       logger5.info("Model loaded successfully", {
         backend: this._backend,
@@ -3214,7 +3262,7 @@ var Wav2ArkitCpuInference = class {
             blendshapes.push(symmetrizeBlendshapes(remapped));
           }
           logger5.trace("Inference completed", {
-            inferenceTimeMs: Math.round(inferenceTimeMs * 100) / 100,
+            inferenceTimeMs: Math.round(inferenceTimeMs),
             numFrames,
             inputSamples
           });
@@ -3280,9 +3328,10 @@ function createLipSync(config) {
     });
   }
   if (useCpu) {
-    logger6.info("Creating Wav2ArkitCpuInference (1.8MB, WASM)");
+    logger6.info("Creating Wav2ArkitCpuInference (WASM)");
     return new Wav2ArkitCpuInference({
-      modelUrl: config.cpuModelUrl
+      modelUrl: config.cpuModelUrl,
+      modelDataUrl: config.cpuModelDataUrl
     });
   }
   const gpuInstance = new Wav2Vec2Inference({
@@ -3309,6 +3358,9 @@ var LipSyncWithFallback = class {
   get isLoaded() {
     return this.implementation.isLoaded;
   }
+  get chunkSamples() {
+    return this.implementation.chunkSamples;
+  }
   async load() {
     try {
       return await this.implementation.load();
@@ -3321,7 +3373,8 @@ var LipSyncWithFallback = class {
       } catch {
       }
       this.implementation = new Wav2ArkitCpuInference({
-        modelUrl: this.config.cpuModelUrl
+        modelUrl: this.config.cpuModelUrl,
+        modelDataUrl: this.config.cpuModelDataUrl
       });
       this.hasFallenBack = true;
       logger6.info("Fallback to Wav2ArkitCpuInference successful");
@@ -3351,8 +3404,6 @@ var SileroVADInference = class {
     // Pre-speech buffer for capturing beginning of speech
     this.preSpeechBuffer = [];
     this.wasSpeaking = false;
-    // Cached sample rate tensor (int64 scalar, never changes per instance)
-    this.srTensor = null;
     const sampleRate = config.sampleRate ?? 16e3;
     if (sampleRate !== 8e3 && sampleRate !== 16e3) {
       throw new Error("Silero VAD only supports 8000 or 16000 Hz sample rates");
@@ -3483,24 +3534,6 @@ var SileroVADInference = class {
     this.context = new Float32Array(this.contextSize);
     this.preSpeechBuffer = [];
     this.wasSpeaking = false;
-    if (!this.srTensor) {
-      try {
-        this.srTensor = new this.ort.Tensor(
-          "int64",
-          new BigInt64Array([BigInt(this.config.sampleRate)]),
-          []
-        );
-      } catch (e) {
-        logger7.warn("BigInt64Array not available, using bigint array fallback", {
-          error: e instanceof Error ? e.message : String(e)
-        });
-        this.srTensor = new this.ort.Tensor(
-          "int64",
-          [BigInt(this.config.sampleRate)],
-          []
-        );
-      }
-    }
   }
   /**
    * Process a single audio chunk
@@ -3632,7 +3665,20 @@ var SileroVADInference = class {
           inputBuffer.set(audioChunkCopy, this.contextSize);
           const inputBufferCopy = new Float32Array(inputBuffer);
           const inputTensor = new this.ort.Tensor("float32", inputBufferCopy, [1, inputSize]);
-          const srTensor = this.srTensor;
+          let srTensor;
+          try {
+            srTensor = new this.ort.Tensor(
+              "int64",
+              new BigInt64Array([BigInt(this.config.sampleRate)]),
+              []
+            );
+          } catch {
+            srTensor = new this.ort.Tensor(
+              "int64",
+              [BigInt(this.config.sampleRate)],
+              []
+            );
+          }
           const stateCopy = new Float32Array(this.state.data);
           const stateTensor = new this.ort.Tensor("float32", stateCopy, this.state.dims);
           const feeds = {
@@ -3721,7 +3767,6 @@ var SileroVADInference = class {
       this.session = null;
     }
     this.state = null;
-    this.srTensor = null;
   }
 };
 /**
@@ -6534,8 +6579,8 @@ async function nukeBrowserCaches(preventRecreation = false) {
       totalDeleted: deletedCount
     });
     if (preventRecreation) {
-      const { env: env2 } = await import("./transformers.web-ALDLCPHT.mjs");
-      env2.useBrowserCache = false;
+      const { env } = await import("./transformers.web-MHLR33H6.mjs");
+      env.useBrowserCache = false;
       logger12.warn("Browser cache creation disabled (env.useBrowserCache = false)");
     }
     return deletedCount;