tts: trim leading silence per chunk via Web Audio decode to skip F5 phantom-phoneme artifacts

public/src/tts.js

@@ -131,10 +131,48 @@ function splitTextIntoChunks(text) {
 	return chunks;
 }
 
+// Shared AudioContext for decoding-only operations (silence detection).
+// Lazily created so we don't hit iOS Safari's "must come from gesture" restriction
+// unless we actually need it.
+let _decodeCtx = null;
+function getDecodeCtx() {
+	if (!_decodeCtx) {
+		const Ctx = window.AudioContext || window.webkitAudioContext;
+		if (!Ctx) return null;
+		_decodeCtx = new Ctx();
+	}
+	return _decodeCtx;
+}
+
+// Detect where speech actually begins in an audio blob. Scans the first 1.5s of
+// decoded PCM for the first sample whose absolute amplitude crosses `threshold`,
+// then returns that time minus a 20ms safety margin so we don't clip the first
+// phoneme. Returns 0 on any decode failure (degrade gracefully).
+async function detectSpeechStart(blob, threshold = 0.05, safetyMs = 20) {
+	try {
+		const ctx = getDecodeCtx();
+		if (!ctx) return 0;
+		const arrayBuffer = await blob.arrayBuffer();
+		const audioBuffer = await ctx.decodeAudioData(arrayBuffer.slice(0));
+		const data = audioBuffer.getChannelData(0);
+		const sampleRate = audioBuffer.sampleRate;
+		const scanSamples = Math.min(data.length, Math.floor(sampleRate * 1.5));
+		for (let i = 0; i < scanSamples; i++) {
+			if (Math.abs(data[i]) > threshold) {
+				return Math.max(0, i / sampleRate - safetyMs / 1000);
+			}
+		}
+		return 0;
+	} catch (e) {
+		console.warn("detectSpeechStart failed:", e);
+		return 0;
+	}
+}
+
 // Pipelined playlist player: fetches all chunks in parallel, plays each as soon
 // as ready, chains via 'ended'. Exposes Audio-compatible API so downstream UI
-// (pause/play, download, progress) works unchanged. Bonus: fixes Safari MP3-concat
-// bug by playing each chunk as its own <audio> element.
+// (pause/play, download, progress) works unchanged. Trims leading silence and
+// F5 phantom-phoneme noise per chunk via Web Audio decode + currentTime skip.
 class PlaylistAudio {
 	constructor(chunks, voice, model, rate) {
 		this._chunks = chunks;
@@ -157,11 +195,26 @@ class PlaylistAudio {
 		try {
 			const blob = await fetchTTSChunk(text, voice, model);
 			this._blobs[idx] = blob;
+
+			// Detect where actual speech starts so we can skip past F5's leading
+			// silence (~300ms) and the phantom-phoneme noise burst F5 emits before
+			// the real first word of each chunk. Without this, users hear "random
+			// chars" at the start of each chunk's audio.
+			const leadOffset = await detectSpeechStart(blob);
+
 			const audio = new Audio(URL.createObjectURL(blob));
 			audio.playbackRate = this._rate;
+			audio.preload = "auto";
+			if (leadOffset > 0) {
+				const setStart = () => {
+					try { audio.currentTime = leadOffset; } catch (_e) {}
+				};
+				if (audio.readyState >= 1) setStart();
+				else audio.addEventListener("loadedmetadata", setStart, { once: true });
+			}
 			audio.addEventListener('ended', () => this._advance(idx));
 			this._audios[idx] = audio;
-			console.log(`PlaylistAudio: chunk ${idx + 1}/${this._chunks.length} ready`);
+			console.log(`PlaylistAudio: chunk ${idx + 1}/${this._chunks.length} ready (skip leading ${(leadOffset * 1000).toFixed(0)}ms)`);
 			if (idx === this._idx && this._wantPlay) {
 				this._wantPlay = false;
 				this._playCurrent();