@tryhamster/gerbil 1.0.0-rc.10 → 1.0.0-rc.11

package/dist/browser/index.js

@@ -1290,6 +1290,126 @@ const TTS_MODELS = {
 		voices: SUPERTONIC_BROWSER_VOICES
 	}
 };
+const TTS_WORKER_CODE = `
+  // TTS Worker - runs in separate thread, loads from CDN
+  import { pipeline, env } from "https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.8.1";
+  
+  // Configure environment
+  env.useBrowserCache = true;
+  env.allowLocalModels = false;
+  
+  let ttsInstance = null;
+  let modelType = null; // "supertonic" or "kokoro"
+  let voiceEmbeddings = new Map();
+  let kokoroTTS = null;
+  
+  self.onmessage = async (e) => {
+    const { type, payload } = e.data;
+    
+    if (type === "load") {
+      try {
+        const { modelId, repo, voices } = payload;
+        modelType = modelId === "supertonic-66m" ? "supertonic" : "kokoro";
+        
+        if (modelType === "supertonic") {
+          // Load Supertonic using transformers.js pipeline
+          ttsInstance = await pipeline("text-to-speech", repo, {
+            device: "webgpu",
+            progress_callback: (progress) => {
+              self.postMessage({ type: "progress", payload: progress });
+            },
+          });
+          
+          // Load voice embeddings
+          for (const voice of voices) {
+            try {
+              const voiceUrl = "https://huggingface.co/" + repo + "/resolve/main/voices/" + voice.id + ".bin";
+              const response = await fetch(voiceUrl);
+              if (response.ok) {
+                const buffer = await response.arrayBuffer();
+                voiceEmbeddings.set(voice.id, new Float32Array(buffer));
+              }
+            } catch (err) {
+              console.warn("Failed to load voice:", voice.id, err);
+            }
+          }
+          
+          // Warmup
+          try {
+            await ttsInstance("Hello", {
+              speaker_embeddings: new Float32Array(1 * 101 * 128),
+              num_inference_steps: 1,
+              speed: 1.0,
+            });
+          } catch (e) {
+            console.warn("Warmup failed:", e);
+          }
+        } else {
+          // Load Kokoro using kokoro-js from CDN
+          const kokoroModule = await import("https://cdn.jsdelivr.net/npm/kokoro-js@1.2.1/dist/kokoro.web.min.js");
+          const { KokoroTTS } = kokoroModule;
+          
+          kokoroTTS = await KokoroTTS.from_pretrained(repo, {
+            dtype: "fp32",
+            progress_callback: (progress) => {
+              self.postMessage({ type: "progress", payload: progress });
+            },
+          });
+        }
+        
+        self.postMessage({ type: "ready" });
+      } catch (err) {
+        self.postMessage({ type: "error", payload: err.message || String(err) });
+      }
+    }
+    
+    if (type === "generate") {
+      try {
+        const { text, voice, speed } = payload;
+        let audio, sampleRate;
+        
+        if (modelType === "supertonic") {
+          let embedding = voiceEmbeddings.get(voice);
+          if (!embedding) {
+            embedding = new Float32Array(101 * 128).fill(0.1);
+          }
+          
+          const result = await ttsInstance(text, {
+            speaker_embeddings: embedding,
+            speed: speed || 1.0,
+          });
+          
+          audio = result.audio;
+          sampleRate = result.sampling_rate;
+        } else {
+          const result = await kokoroTTS.generate(text, {
+            voice: voice,
+            speed: speed || 1.0,
+          });
+          
+          audio = result.audio;
+          sampleRate = result.sampling_rate;
+        }
+        
+        // Transfer audio data back
+        self.postMessage(
+          { type: "audio", payload: { audio: audio, sampleRate: sampleRate } },
+          [audio.buffer]
+        );
+      } catch (err) {
+        self.postMessage({ type: "error", payload: err.message || String(err) });
+      }
+    }
+  };
+`;
+/** Create TTS worker instance */
+function createTTSWorker() {
+	const blob = new Blob([TTS_WORKER_CODE], { type: "application/javascript" });
+	const url = URL.createObjectURL(blob);
+	const worker = new Worker(url, { type: "module" });
+	URL.revokeObjectURL(url);
+	return worker;
+}
 /**
 * React hook for text-to-speech with Web Audio API playback
 *
@@ -1339,17 +1459,17 @@ function useSpeech(options = {}) {
 	const [shouldLoad, setShouldLoad] = useState(autoLoad);
 	const [currentVoice, setCurrentVoice] = useState(defaultVoice);
 	const [currentSpeed, setCurrentSpeed] = useState(defaultSpeed);
-	const ttsRef = useRef(null);
-	const voiceEmbeddingsRef = useRef(/* @__PURE__ */ new Map());
+	const workerRef = useRef(null);
 	const audioContextRef = useRef(null);
 	const sourceNodeRef = useRef(null);
 	const mountedRef = useRef(true);
 	const modelIdRef = useRef(modelId);
+	const pendingSpeakRef = useRef(null);
 	const listVoices = useCallback(() => {
 		return modelConfig.voices;
 	}, [modelConfig.voices]);
 	const load = useCallback(() => {
-		if (ttsRef.current || isLoading) return;
+		if (workerRef.current || isLoading) return;
 		setIsLoading(true);
 		setShouldLoad(true);
 	}, [isLoading]);
@@ -1357,87 +1477,48 @@ function useSpeech(options = {}) {
 		if (!shouldLoad) return;
 		mountedRef.current = true;
 		modelIdRef.current = modelId;
-		const initTTS = async () => {
-			try {
-				const isSupertonic = modelId === "supertonic-66m";
-				const config = TTS_MODELS[modelId];
-				setLoadingProgress({
-					status: "loading",
-					message: `Loading ${isSupertonic ? "Supertonic" : "Kokoro"} TTS...`
-				});
-				if (isSupertonic) {
-					const { pipeline, env } = await import("../transformers.web-u34VxRFM.js");
-					if (env.backends?.onnx?.wasm) env.backends.onnx.wasm.wasmPaths = "https://cdn.jsdelivr.net/npm/onnxruntime-web@1.21.0/dist/";
-					const tts = await pipeline("text-to-speech", config.repo, {
-						device: "webgpu",
-						progress_callback: (progress) => {
-							if (!mountedRef.current) return;
-							if (progress.status === "progress" && progress.file) setLoadingProgress({
-								status: "downloading",
-								file: progress.file,
-								progress: Math.round(progress.progress || 0)
-							});
-						}
-					});
-					if (!mountedRef.current) return;
-					const voicesUrl = `https://huggingface.co/${config.repo}/resolve/main/voices/`;
-					const embeddingsMap = /* @__PURE__ */ new Map();
-					await Promise.all(config.voices.map(async (voice) => {
-						try {
-							const response = await fetch(`${voicesUrl}${voice.id}.bin`);
-							if (response.ok) {
-								const buffer = await response.arrayBuffer();
-								embeddingsMap.set(voice.id, new Float32Array(buffer));
-							}
-						} catch (e) {
-							console.warn(`Failed to load voice embedding for ${voice.id}:`, e);
-						}
-					}));
-					if (!mountedRef.current) return;
-					try {
-						await tts("Hello", {
-							speaker_embeddings: new Float32Array(12928),
-							num_inference_steps: 1,
-							speed: 1
-						});
-					} catch (e) {
-						console.warn("Supertonic warmup failed:", e);
-					}
-					voiceEmbeddingsRef.current = embeddingsMap;
-					ttsRef.current = {
-						type: "supertonic",
-						pipeline: tts,
-						config
-					};
-				} else {
-					const { KokoroTTS } = await import("../kokoro-CMOGDSgT.js");
-					const tts = await KokoroTTS.from_pretrained(config.repo, {
-						dtype: "fp32",
-						progress_callback: (progress) => {
-							if (!mountedRef.current) return;
-							if (progress.status === "progress" && progress.file) setLoadingProgress({
-								status: "downloading",
-								file: progress.file,
-								progress: Math.round(progress.progress || 0)
-							});
-						}
-					});
-					if (!mountedRef.current) return;
-					ttsRef.current = {
-						type: "kokoro",
-						instance: tts,
-						config
-					};
-				}
+		const config = TTS_MODELS[modelId];
+		setLoadingProgress({
+			status: "loading",
+			message: `Loading ${modelId === "supertonic-66m" ? "Supertonic" : "Kokoro"} TTS...`
+		});
+		const worker = createTTSWorker();
+		workerRef.current = worker;
+		worker.onmessage = (e) => {
+			if (!mountedRef.current) return;
+			const { type, payload } = e.data;
+			if (type === "progress" && payload.status === "progress" && payload.file) setLoadingProgress({
+				status: "downloading",
+				file: payload.file,
+				progress: Math.round(payload.progress || 0)
+			});
+			if (type === "ready") {
 				setIsLoading(false);
 				setIsReady(true);
 				setLoadingProgress({ status: "ready" });
 				onReady?.();
-			} catch (err) {
-				if (!mountedRef.current) return;
-				const errorMsg = err instanceof Error ? err.message : String(err);
+				if (pendingSpeakRef.current) {
+					const { text, voice, speed } = pendingSpeakRef.current;
+					pendingSpeakRef.current = null;
+					worker.postMessage({
+						type: "generate",
+						payload: {
+							text,
+							voice,
+							speed
+						}
+					});
+				}
+			}
+			if (type === "audio") {
+				const { audio, sampleRate } = payload;
+				playAudioData(audio, sampleRate);
+			}
+			if (type === "error") {
+				const errorMsg = payload;
 				setError(errorMsg);
 				setIsLoading(false);
+				setIsSpeaking(false);
 				setLoadingProgress({
 					status: "error",
 					error: errorMsg
@@ -1445,9 +1526,29 @@ function useSpeech(options = {}) {
 				onError?.(errorMsg);
 			}
 		};
-		initTTS();
+		worker.onerror = (err) => {
+			if (!mountedRef.current) return;
+			const errorMsg = err.message || "Worker error";
+			setError(errorMsg);
+			setIsLoading(false);
+			setLoadingProgress({
+				status: "error",
+				error: errorMsg
+			});
+			onError?.(errorMsg);
+		};
+		worker.postMessage({
+			type: "load",
+			payload: {
+				modelId,
+				repo: config.repo,
+				voices: config.voices
+			}
+		});
 		return () => {
 			mountedRef.current = false;
+			worker.terminate();
+			workerRef.current = null;
 		};
 	}, [
 		shouldLoad,
@@ -1455,6 +1556,30 @@ function useSpeech(options = {}) {
 		onReady,
 		onError
 	]);
+	const playAudioData = useCallback(async (audio, sampleRate) => {
+		try {
+			if (!audioContextRef.current || audioContextRef.current.state === "closed") audioContextRef.current = new AudioContext({ sampleRate });
+			const ctx = audioContextRef.current;
+			if (ctx.state === "suspended") await ctx.resume();
+			const audioBuffer = ctx.createBuffer(1, audio.length, sampleRate);
+			audioBuffer.copyToChannel(new Float32Array(audio), 0);
+			const sourceNode = ctx.createBufferSource();
+			sourceNode.buffer = audioBuffer;
+			sourceNode.connect(ctx.destination);
+			sourceNodeRef.current = sourceNode;
+			sourceNode.onended = () => {
+				if (!mountedRef.current) return;
+				setIsSpeaking(false);
+				onEnd?.();
+			};
+			sourceNode.start();
+		} catch (err) {
+			setIsSpeaking(false);
+			const errorMsg = err instanceof Error ? err.message : String(err);
+			setError(errorMsg);
+			onError?.(errorMsg);
+		}
+	}, [onEnd, onError]);
 	useEffect(() => {
 		return () => {
 			try {
@@ -1469,89 +1594,46 @@ function useSpeech(options = {}) {
 		speak: useCallback(async (text, opts) => {
 			const voice = opts?.voice || currentVoice;
 			const speed = opts?.speed || currentSpeed;
-			if (!ttsRef.current) {
+			if (!modelConfig.voices.find((v) => v.id === voice)) {
+				const errorMsg = `Voice "${voice}" not found. Should be one of: ${modelConfig.voices.map((v) => v.id).join(", ")}.`;
+				setError(errorMsg);
+				onError?.(errorMsg);
+				return;
+			}
+			if (!workerRef.current) {
+				pendingSpeakRef.current = {
+					text,
+					voice,
+					speed
+				};
 				load();
 				return;
 			}
-			try {
-				setIsSpeaking(true);
-				onStart?.();
-				let audioData;
-				let sampleRate;
-				const ttsBackend = ttsRef.current;
-				if (ttsBackend.type === "supertonic") {
-					const config = ttsBackend.config;
-					if (!config.voices.find((v) => v.id === voice)) {
-						const validVoices = config.voices.map((v) => v.id).join(", ");
-						throw new Error(`Voice "${voice}" not found. Should be one of: ${validVoices}.`);
-					}
-					let speakerEmbedding = voiceEmbeddingsRef.current.get(voice);
-					if (!speakerEmbedding) try {
-						const voiceUrl = `https://huggingface.co/${config.repo}/resolve/main/voices/${voice}.bin`;
-						const response = await fetch(voiceUrl);
-						if (response.ok) {
-							const buffer = await response.arrayBuffer();
-							speakerEmbedding = new Float32Array(buffer);
-							voiceEmbeddingsRef.current.set(voice, speakerEmbedding);
-						} else throw new Error(`Failed to load voice: ${response.status}`);
-					} catch {
-						speakerEmbedding = new Float32Array(12928).fill(.1);
-						voiceEmbeddingsRef.current.set(voice, speakerEmbedding);
-					}
-					const result = await ttsBackend.pipeline(text, {
-						speaker_embeddings: speakerEmbedding,
-						speed
-					});
-					audioData = result.audio;
-					sampleRate = result.sampling_rate;
-				} else {
-					const config = ttsBackend.config;
-					if (!config.voices.find((v) => v.id === voice)) {
-						const validVoices = config.voices.map((v) => v.id).join(", ");
-						throw new Error(`Voice "${voice}" not found. Should be one of: ${validVoices}.`);
-					}
-					const result = await ttsBackend.instance.generate(text, {
-						voice,
-						speed
-					});
-					audioData = result.audio;
-					sampleRate = result.sampling_rate;
-				}
-				if (!mountedRef.current) return;
-				if (!audioContextRef.current || audioContextRef.current.state === "closed") audioContextRef.current = new AudioContext();
-				const audioContext = audioContextRef.current;
-				if (audioContext.state === "suspended") await audioContext.resume();
-				const audioBuffer = audioContext.createBuffer(1, audioData.length, sampleRate);
-				const channelData = new Float32Array(audioData);
-				audioBuffer.copyToChannel(channelData, 0);
-				if (sourceNodeRef.current) {
-					sourceNodeRef.current.stop();
-					sourceNodeRef.current.disconnect();
-				}
-				const sourceNode = audioContext.createBufferSource();
-				sourceNode.buffer = audioBuffer;
-				sourceNode.connect(audioContext.destination);
-				sourceNode.onended = () => {
-					if (mountedRef.current) {
-						setIsSpeaking(false);
-						onEnd?.();
-					}
+			if (!isReady) {
+				pendingSpeakRef.current = {
+					text,
+					voice,
+					speed
 				};
-				sourceNodeRef.current = sourceNode;
-				sourceNode.start();
-			} catch (err) {
-				if (!mountedRef.current) return;
-				const errorMsg = err instanceof Error ? err.message : String(err);
-				setError(errorMsg);
-				setIsSpeaking(false);
-				onError?.(errorMsg);
+				return;
 			}
+			setIsSpeaking(true);
+			onStart?.();
+			workerRef.current.postMessage({
+				type: "generate",
+				payload: {
+					text,
+					voice,
+					speed
+				}
+			});
 		}, [
 			currentVoice,
 			currentSpeed,
+			modelConfig.voices,
 			load,
+			isReady,
 			onStart,
-			onEnd,
 			onError
 		]),
 		stop: useCallback(() => {
@@ -1674,6 +1756,61 @@ function createAudioPlayer(sampleRate = 24e3) {
 		isPlaying: () => isActive
 	};
 }
+const STT_WORKER_CODE = `
+  // STT Worker - runs in separate thread, loads from CDN
+  import { pipeline, env } from "https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.8.1";
+  
+  // Configure environment
+  env.useBrowserCache = true;
+  env.allowLocalModels = false;
+  
+  let sttPipeline = null;
+  
+  self.onmessage = async (e) => {
+    const { type, payload } = e.data;
+    
+    if (type === "load") {
+      try {
+        const { model } = payload;
+        
+        // Load Whisper model
+        sttPipeline = await pipeline("automatic-speech-recognition", model, {
+          device: "webgpu",
+          progress_callback: (progress) => {
+            self.postMessage({ type: "progress", payload: progress });
+          },
+        });
+        
+        self.postMessage({ type: "ready" });
+      } catch (err) {
+        self.postMessage({ type: "error", payload: err.message || String(err) });
+      }
+    }
+    
+    if (type === "transcribe") {
+      try {
+        const { audio } = payload;
+        
+        // Run transcription
+        const result = await sttPipeline(audio, {
+          return_timestamps: false,
+        });
+        
+        self.postMessage({ type: "transcript", payload: result.text || "" });
+      } catch (err) {
+        self.postMessage({ type: "error", payload: err.message || String(err) });
+      }
+    }
+  };
+`;
+/** Create STT worker instance */
+function createSTTWorker() {
+	const blob = new Blob([STT_WORKER_CODE], { type: "application/javascript" });
+	const url = URL.createObjectURL(blob);
+	const worker = new Worker(url, { type: "module" });
+	URL.revokeObjectURL(url);
+	return worker;
+}
 /**
 * React hook for voice input with browser microphone
 *
@@ -1731,7 +1868,7 @@ function useVoiceInput(options = {}) {
 	const [chunkCount, setChunkCount] = useState(0);
 	const [error, setError] = useState(null);
 	const [shouldLoad, setShouldLoad] = useState(autoLoad);
-	const sttRef = useRef(null);
+	const workerRef = useRef(null);
 	const mediaRecorderRef = useRef(null);
 	const audioChunksRef = useRef([]);
 	const streamRef = useRef(null);
@@ -1739,49 +1876,66 @@ function useVoiceInput(options = {}) {
 	const streamingIntervalRef = useRef(null);
 	const pendingChunksRef = useRef([]);
 	const fullTranscriptRef = useRef("");
+	const transcribeResolveRef = useRef(null);
+	const transcribeRejectRef = useRef(null);
+	const resolveSTTModel = (modelId) => {
+		return {
+			"whisper-tiny": "onnx-community/whisper-tiny",
+			"whisper-tiny.en": "onnx-community/whisper-tiny.en",
+			"whisper-base": "onnx-community/whisper-base",
+			"whisper-base.en": "onnx-community/whisper-base.en",
+			"whisper-small": "onnx-community/whisper-small",
+			"whisper-small.en": "onnx-community/whisper-small.en"
+		}[modelId] || modelId;
+	};
 	useEffect(() => {
 		if (!shouldLoad || isReady) return;
-		let cancelled = false;
-		const loadModel = async () => {
-			try {
-				setIsLoading(true);
-				setLoadingProgress({
-					status: "loading",
-					message: "Loading STT model..."
-				});
-				onProgress?.({
-					status: "loading",
-					message: "Loading STT model..."
-				});
-				const { WhisperSTT } = await import("../stt-Dne6SENv.js");
-				if (cancelled || !mountedRef.current) return;
-				const stt = new WhisperSTT(model);
-				await stt.load({ onProgress: (p) => {
-					if (!mountedRef.current) return;
-					const progress = {
-						status: p.progress !== void 0 ? "downloading" : "loading",
-						message: p.status,
-						progress: p.progress,
-						file: p.file
-					};
-					setLoadingProgress(progress);
-					onProgress?.(progress);
-				} });
-				if (cancelled || !mountedRef.current) {
-					stt.dispose();
-					return;
-				}
-				sttRef.current = stt;
+		mountedRef.current = true;
+		setIsLoading(true);
+		setLoadingProgress({
+			status: "loading",
+			message: "Loading STT model..."
+		});
+		onProgress?.({
+			status: "loading",
+			message: "Loading STT model..."
+		});
+		const worker = createSTTWorker();
+		workerRef.current = worker;
+		worker.onmessage = (e) => {
+			if (!mountedRef.current) return;
+			const { type, payload } = e.data;
+			if (type === "progress") {
+				const progress = {
+					status: payload.progress !== void 0 ? "downloading" : "loading",
+					message: payload.status,
+					progress: payload.progress,
+					file: payload.file
+				};
+				setLoadingProgress(progress);
+				onProgress?.(progress);
+			}
+			if (type === "ready") {
 				setIsReady(true);
 				setIsLoading(false);
 				setLoadingProgress({ status: "ready" });
 				onProgress?.({ status: "ready" });
 				onReady?.();
-			} catch (e) {
-				if (!mountedRef.current) return;
-				const errMsg = e.message || "Failed to load STT model";
+			}
+			if (type === "transcript") {
+				const text = payload;
+				setIsTranscribing(false);
+				if (transcribeResolveRef.current) {
+					transcribeResolveRef.current(text);
+					transcribeResolveRef.current = null;
+					transcribeRejectRef.current = null;
+				}
+			}
+			if (type === "error") {
+				const errMsg = payload;
 				setError(errMsg);
 				setIsLoading(false);
+				setIsTranscribing(false);
 				setLoadingProgress({
 					status: "error",
 					message: errMsg
@@ -1791,11 +1945,36 @@ function useVoiceInput(options = {}) {
 					message: errMsg
 				});
 				onError?.(errMsg);
+				if (transcribeRejectRef.current) {
+					transcribeRejectRef.current(new Error(errMsg));
+					transcribeResolveRef.current = null;
+					transcribeRejectRef.current = null;
+				}
 			}
 		};
-		loadModel();
+		worker.onerror = (err) => {
+			if (!mountedRef.current) return;
+			const errMsg = err.message || "Worker error";
+			setError(errMsg);
+			setIsLoading(false);
+			setLoadingProgress({
+				status: "error",
+				message: errMsg
+			});
+			onProgress?.({
+				status: "error",
+				message: errMsg
+			});
+			onError?.(errMsg);
+		};
+		worker.postMessage({
+			type: "load",
+			payload: { model: resolveSTTModel(model) }
+		});
 		return () => {
-			cancelled = true;
+			mountedRef.current = false;
+			worker.terminate();
+			workerRef.current = null;
 		};
 	}, [
 		shouldLoad,
@@ -1809,7 +1988,10 @@ function useVoiceInput(options = {}) {
 		mountedRef.current = true;
 		return () => {
 			mountedRef.current = false;
-			if (sttRef.current) sttRef.current.dispose();
+			if (workerRef.current) {
+				workerRef.current.terminate();
+				workerRef.current = null;
+			}
 			if (streamRef.current) for (const track of streamRef.current.getTracks()) track.stop();
 		};
 	}, []);
@@ -1843,27 +2025,38 @@ function useVoiceInput(options = {}) {
 		return new Float32Array(channelData);
 	}, []);
 	const transcribe = useCallback(async (audio) => {
-		if (!sttRef.current) {
+		if (!workerRef.current) {
 			if (!shouldLoad) {
 				setShouldLoad(true);
 				throw new Error("STT model not loaded. Loading now, please try again.");
 			}
 			throw new Error("STT model not loaded");
 		}
+		if (!isReady) throw new Error("STT model still loading");
 		setIsTranscribing(true);
-		try {
-			let text = (await sttRef.current.transcribe(audio)).text.trim();
-			if (text === "[BLANK_AUDIO]" || text === "(blank audio)" || text === "[BLANK AUDIO]") text = "";
-			setTranscript(text);
-			onTranscript?.(text);
-			return text;
-		} finally {
-			if (mountedRef.current) setIsTranscribing(false);
-		}
-	}, [shouldLoad, onTranscript]);
+		return new Promise((resolve, reject) => {
+			transcribeResolveRef.current = (text) => {
+				let filtered = text.trim();
+				if (filtered === "[BLANK_AUDIO]" || filtered === "(blank audio)" || filtered === "[BLANK AUDIO]") filtered = "";
+				setTranscript(filtered);
+				onTranscript?.(filtered);
+				resolve(filtered);
+			};
+			transcribeRejectRef.current = reject;
+			const audioArray = new Float32Array(audio);
+			workerRef.current.postMessage({
+				type: "transcribe",
+				payload: { audio: audioArray }
+			}, [audioArray.buffer]);
+		});
+	}, [
+		shouldLoad,
+		isReady,
+		onTranscript
+	]);
 	const processedSamplesRef = useRef(0);
 	const transcribeChunk = useCallback(async (chunkIdx) => {
-		if (!sttRef.current || audioChunksRef.current.length === 0) return "";
+		if (!workerRef.current || !isReady || audioChunksRef.current.length === 0) return "";
 		try {
 			const audioData = await blobToFloat32(new Blob(audioChunksRef.current, { type: "audio/webm" }));
 			const newSamplesStart = processedSamplesRef.current;
@@ -1871,8 +2064,7 @@ function useVoiceInput(options = {}) {
 			if (totalSamples - newSamplesStart < 8e3) return "";
 			const newAudio = audioData.slice(newSamplesStart);
 			processedSamplesRef.current = totalSamples;
-			let text = (await sttRef.current.transcribe(newAudio)).text.trim();
-			if (text === "[BLANK_AUDIO]" || text === "(blank audio)" || text === "[BLANK AUDIO]") text = "";
+			const text = await transcribe(newAudio);
 			if (text && mountedRef.current) {
 				setStreamingChunk(text);
 				onChunk?.(text, chunkIdx);
@@ -1881,38 +2073,30 @@ function useVoiceInput(options = {}) {
 		} catch {
 			return "";
 		}
-	}, [blobToFloat32, onChunk]);
+	}, [
+		blobToFloat32,
+		isReady,
+		transcribe,
+		onChunk
+	]);
 	return {
 		startRecording: useCallback(async () => {
 			if (isRecording) return;
 			try {
-				if (streaming && !sttRef.current) {
+				if (streaming && !isReady) {
 					if (!shouldLoad) setShouldLoad(true);
-					setIsLoading(true);
-					const { WhisperSTT } = await import("../stt-Dne6SENv.js");
-					const stt = new WhisperSTT(model);
-					await stt.load({ onProgress: (p) => {
-						if (mountedRef.current) {
-							const progress = {
-								status: p.status === "downloading" ? "downloading" : p.status === "ready" ? "ready" : "loading",
-								message: p.status,
-								progress: p.progress,
-								file: p.file
-							};
-							setLoadingProgress(progress);
-							onProgress?.(progress);
-						}
-					} });
-					if (!mountedRef.current) {
-						stt.dispose();
-						return;
-					}
-					sttRef.current = stt;
-					setIsReady(true);
-					setIsLoading(false);
-					setLoadingProgress({ status: "ready" });
-					onProgress?.({ status: "ready" });
-					onReady?.();
+					await new Promise((resolve, reject) => {
+						const checkReady = setInterval(() => {
+							if (isReady && workerRef.current) {
+								clearInterval(checkReady);
+								resolve();
+							}
+						}, 100);
+						setTimeout(() => {
+							clearInterval(checkReady);
+							reject(/* @__PURE__ */ new Error("Timeout waiting for STT model"));
+						}, 6e4);
+					});
 				}
 				const stream = await navigator.mediaDevices.getUserMedia({ audio: {
 					sampleRate: 16e3,
@@ -1939,7 +2123,7 @@ function useVoiceInput(options = {}) {
 				mediaRecorder.start(100);
 				setIsRecording(true);
 				setError(null);
-				if (streaming && sttRef.current) {
+				if (streaming && isReady && workerRef.current) {
 					let chunkIdx = 0;
 					let shouldContinue = true;
 					const processNextChunk = async () => {
@@ -2029,11 +2213,11 @@ function useVoiceInput(options = {}) {
 					}
 					const audioBlob = new Blob(audioChunksRef.current, { type: "audio/webm" });
 					try {
-						if (!sttRef.current) {
+						if (!isReady || !workerRef.current) {
 							if (!shouldLoad) setShouldLoad(true);
 							await new Promise((res, rej) => {
 								const checkReady = setInterval(() => {
-									if (sttRef.current) {
+									if (isReady && workerRef.current) {
 										clearInterval(checkReady);
 										res();
 									}
@@ -2164,6 +2348,16 @@ function useVoiceChat(options = {}) {
 	const isListening = stage === "listening";
 	const isProcessing = stage === "transcribing" || stage === "thinking";
 	const isSpeaking = stage === "speaking";
+	const resolveSTTModel = (modelId) => {
+		return {
+			"whisper-tiny": "onnx-community/whisper-tiny",
+			"whisper-tiny.en": "onnx-community/whisper-tiny.en",
+			"whisper-base": "onnx-community/whisper-base",
+			"whisper-base.en": "onnx-community/whisper-base.en",
+			"whisper-small": "onnx-community/whisper-small",
+			"whisper-small.en": "onnx-community/whisper-small.en"
+		}[modelId] || modelId;
+	};
 	useEffect(() => {
 		if (!shouldLoad || isReady) return;
 		let cancelled = false;
@@ -2172,18 +2366,29 @@ function useVoiceChat(options = {}) {
 				setIsLoading(true);
 				setError(null);
 				setLoadingMessage("Loading speech recognition (Whisper)...");
-				const { WhisperSTT } = await import("../stt-Dne6SENv.js");
-				if (cancelled || !mountedRef.current) return;
-				const stt = new WhisperSTT(sttModel);
-				await stt.load({ onProgress: (p) => {
-					if (!mountedRef.current) return;
-					setLoadingMessage(p.status || "Loading STT...");
-				} });
+				const sttWorker = createSTTWorker();
+				if (cancelled || !mountedRef.current) {
+					sttWorker.terminate();
+					return;
+				}
+				await new Promise((resolve, reject) => {
+					sttWorker.onmessage = (e) => {
+						const { type, payload } = e.data;
+						if (type === "ready") resolve();
+						if (type === "error") reject(new Error(payload));
+						if (type === "progress" && mountedRef.current) setLoadingMessage(payload.status || "Loading STT...");
+					};
+					sttWorker.onerror = (e) => reject(new Error(e.message));
+					sttWorker.postMessage({
+						type: "load",
+						payload: { model: resolveSTTModel(sttModel) }
+					});
+				});
 				if (cancelled || !mountedRef.current) {
-					stt.dispose();
+					sttWorker.terminate();
 					return;
 				}
-				sttRef.current = stt;
+				sttRef.current = sttWorker;
 				setLoadingMessage("Loading language model...");
 				const worker = await createGerbilWorker({
 					modelId: llmModel,
@@ -2198,18 +2403,34 @@ function useVoiceChat(options = {}) {
 				}
 				llmWorkerRef.current = worker;
 				setLoadingMessage(`Loading text-to-speech (${ttsModelId === "supertonic-66m" ? "Supertonic" : "Kokoro"})...`);
-				const { createTTS } = await import("../tts-C2FzKuSx.js");
-				if (cancelled || !mountedRef.current) return;
-				const tts = createTTS(ttsModelId);
-				await tts.load({ onProgress: (p) => {
-					if (!mountedRef.current) return;
-					setLoadingMessage(p.status || "Loading TTS...");
-				} });
+				const ttsWorker = createTTSWorker();
+				if (cancelled || !mountedRef.current) {
+					ttsWorker.terminate();
+					return;
+				}
+				const ttsConfig$1 = TTS_MODELS[ttsModelId];
+				await new Promise((resolve, reject) => {
+					ttsWorker.onmessage = (e) => {
+						const { type, payload } = e.data;
+						if (type === "ready") resolve();
+						if (type === "error") reject(new Error(payload));
+						if (type === "progress" && mountedRef.current) setLoadingMessage(payload.status || "Loading TTS...");
+					};
+					ttsWorker.onerror = (e) => reject(new Error(e.message));
+					ttsWorker.postMessage({
+						type: "load",
+						payload: {
+							modelId: ttsModelId,
+							repo: ttsConfig$1.repo,
+							voices: ttsConfig$1.voices
+						}
+					});
+				});
 				if (cancelled || !mountedRef.current) {
-					await tts.dispose();
+					ttsWorker.terminate();
 					return;
 				}
-				ttsRef.current = tts;
+				ttsRef.current = ttsWorker;
 				setIsReady(true);
 				setIsLoading(false);
 				setLoadingMessage("Ready!");
@@ -2238,8 +2459,8 @@ function useVoiceChat(options = {}) {
 		return () => {
 			mountedRef.current = false;
 			llmWorkerRef.current?.terminate();
-			sttRef.current?.dispose();
-			ttsRef.current?.dispose();
+			sttRef.current?.terminate();
+			ttsRef.current?.terminate();
 			if (streamRef.current) for (const track of streamRef.current.getTracks()) track.stop();
 			audioContextRef.current?.close();
 		};
@@ -2345,7 +2566,26 @@ function useVoiceChat(options = {}) {
 					try {
 						setStage("transcribing");
 						const audioData = await blobToFloat32(audioBlob);
-						let userText = (await sttRef.current.transcribe(audioData)).text.trim();
+						let userText = await new Promise((sttResolve, sttReject) => {
+							const handler = (e) => {
+								const { type, payload } = e.data;
+								if (type === "transcript") {
+									sttRef.current?.removeEventListener("message", handler);
+									sttResolve(payload);
+								}
+								if (type === "error") {
+									sttRef.current?.removeEventListener("message", handler);
+									sttReject(new Error(payload));
+								}
+							};
+							sttRef.current?.addEventListener("message", handler);
+							const audioArray = new Float32Array(audioData);
+							sttRef.current?.postMessage({
+								type: "transcribe",
+								payload: { audio: audioArray }
+							}, [audioArray.buffer]);
+						});
+						userText = userText.trim();
 						if (userText === "[BLANK_AUDIO]" || userText === "(blank audio)" || userText === "[BLANK AUDIO]") userText = "";
 						if (cancelledRef.current || !userText) {
 							setStage("idle");
@@ -2395,9 +2635,30 @@ function useVoiceChat(options = {}) {
 						onAssistantSpeak?.(responseText);
 						if (responseText.trim()) {
 							setStage("speaking");
-							const ttsResult = await ttsRef.current.speak(responseText, {
-								voice,
-								speed
+							const ttsResult = await new Promise((ttsResolve, ttsReject) => {
+								const handler = (e) => {
+									const { type, payload } = e.data;
+									if (type === "audio") {
+										ttsRef.current?.removeEventListener("message", handler);
+										ttsResolve({
+											audio: payload.audio,
+											sampleRate: payload.sampleRate
+										});
+									}
+									if (type === "error") {
+										ttsRef.current?.removeEventListener("message", handler);
+										ttsReject(new Error(payload));
+									}
+								};
+								ttsRef.current?.addEventListener("message", handler);
+								ttsRef.current?.postMessage({
+									type: "generate",
+									payload: {
+										text: responseText,
+										voice,
+										speed
+									}
+								});
 							});
 							if (!cancelledRef.current) await playAudioBuffer(ttsResult.audio, ttsResult.sampleRate);
 						}