@omote/core 0.2.0 → 0.2.2

package/dist/index.d.mts

@@ -366,12 +366,13 @@ declare function isSafari(): boolean;
 /**
  * Recommend using CPU-optimized lip sync model (wav2arkit_cpu)
  *
- * Safari (macOS + iOS) has WebGPU JSEP bugs that crash ONNX Runtime,
- * and the 384MB LAM model stack-overflows in WASM mode.
+ * All WebKit browsers (Safari macOS, Safari iOS, Chrome iOS, Firefox iOS)
+ * have ONNX Runtime WebGPU JSEP bugs that crash session creation, and the
+ * 384MB LAM model stack-overflows in WASM mode.
  * The wav2arkit_cpu model (1.8MB) provides identical 52 ARKit blendshape
  * output at 22x real-time on CPU/WASM.
  *
- * @returns true if on Safari (should use CPU lip sync model)
+ * @returns true if on Safari or any iOS browser (should use CPU lip sync model)
  */
 declare function shouldUseCpuLipSync(): boolean;
 /**
@@ -1366,6 +1367,7 @@ declare class SileroVADInference {
     private inferenceQueue;
     private preSpeechBuffer;
     private wasSpeaking;
+    private srTensor;
     constructor(config: SileroVADConfig);
     get backend(): RuntimeBackend | null;
     get isLoaded(): boolean;

package/dist/index.d.ts

@@ -366,12 +366,13 @@ declare function isSafari(): boolean;
 /**
  * Recommend using CPU-optimized lip sync model (wav2arkit_cpu)
  *
- * Safari (macOS + iOS) has WebGPU JSEP bugs that crash ONNX Runtime,
- * and the 384MB LAM model stack-overflows in WASM mode.
+ * All WebKit browsers (Safari macOS, Safari iOS, Chrome iOS, Firefox iOS)
+ * have ONNX Runtime WebGPU JSEP bugs that crash session creation, and the
+ * 384MB LAM model stack-overflows in WASM mode.
  * The wav2arkit_cpu model (1.8MB) provides identical 52 ARKit blendshape
  * output at 22x real-time on CPU/WASM.
  *
- * @returns true if on Safari (should use CPU lip sync model)
+ * @returns true if on Safari or any iOS browser (should use CPU lip sync model)
  */
 declare function shouldUseCpuLipSync(): boolean;
 /**
@@ -1366,6 +1367,7 @@ declare class SileroVADInference {
     private inferenceQueue;
     private preSpeechBuffer;
     private wasSpeaking;
+    private srTensor;
     constructor(config: SileroVADConfig);
     get backend(): RuntimeBackend | null;
     get isLoaded(): boolean;

package/dist/index.js

@@ -28970,7 +28970,7 @@ function isSafari() {
   return /safari/.test(ua) && !/chrome|crios|fxios|chromium|edg/.test(ua);
 }
 function shouldUseCpuLipSync() {
-  return isSafari();
+  return isSafari() || isIOS();
 }
 function isSpeechRecognitionAvailable() {
   if (typeof window === "undefined") return false;
@@ -30138,7 +30138,7 @@ function createLipSync(config) {
     useCpu = false;
     logger6.info("Forcing GPU lip sync model (Wav2Vec2)");
   } else {
-    useCpu = isSafari();
+    useCpu = shouldUseCpuLipSync();
     logger6.info("Auto-detected lip sync model", {
       useCpu,
       isSafari: isSafari()
@@ -30216,6 +30216,8 @@ 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");
@@ -30346,6 +30348,24 @@ 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
@@ -30477,7 +30497,7 @@ 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 = new this.ort.Tensor("int64", BigInt64Array.from([BigInt(this.config.sampleRate)]), []);
+          const srTensor = this.srTensor;
           const stateCopy = new Float32Array(this.state.data);
           const stateTensor = new this.ort.Tensor("float32", stateCopy, this.state.dims);
           const feeds = {
@@ -30566,6 +30586,7 @@ var SileroVADInference = class {
       this.session = null;
     }
     this.state = null;
+    this.srTensor = null;
   }
 };
 /**
@@ -30670,7 +30691,14 @@ async function runInference(audio, state, context) {
   // Create tensors
   const inputTensor = new ort.Tensor('float32', new Float32Array(inputBuffer), [1, inputSize]);
   const stateTensor = new ort.Tensor('float32', new Float32Array(state), [2, 1, 128]);
-  const srTensor = new ort.Tensor('int64', BigInt64Array.from([BigInt(sampleRate)]), []);
+  // Use BigInt64Array constructor (not .from()) for broader compatibility
+  let srTensor;
+  try {
+    srTensor = new ort.Tensor('int64', new BigInt64Array([BigInt(sampleRate)]), []);
+  } catch (e) {
+    // Fallback for environments without BigInt64Array support
+    srTensor = new ort.Tensor('int64', [BigInt(sampleRate)], []);
+  }
 
   const feeds = {
     'input': inputTensor,