From 075db9154a65a5733b0682445c57ffee589738b1 Mon Sep 17 00:00:00 2001
From: "russell@unturf.com" <russell@unturf.com>
Date: Sun, 24 May 2026 10:20:52 -0400
Subject: [PATCH] tts: punctuate unbounded chunks, strip URLs and paths, log
 chunk text

---
 public/src/tts.js | 39 +++++++++++++++++++++++++++++++--------
 1 file changed, 31 insertions(+), 8 deletions(-)

diff --git a/public/src/tts.js b/public/src/tts.js
index 2f7261c..87877d5 100644
--- a/public/src/tts.js
+++ b/public/src/tts.js
@@ -73,28 +73,49 @@ const ttsCache = new Map();
 const TTS_CHUNK_MAX = 2500;       // upper bound per chunk (server-side comfortable)
 const TTS_FIRST_CHUNK_MAX = 500;  // smaller first chunk for low time-to-first-audio
 
+// Strip patterns F5 would pronounce character-by-character (URLs, paths, emails).
+// F5 has no spell-vs-speak heuristic — feeding it https://foo.bar produces audible
+// garbage that users hear as "random chars trying to be pronounced".
+function stripUnspeakable(text) {
+	return text
+		.replace(/https?:\/\/\S+/gi, "")
+		.replace(/\b[\w-]+@[\w.-]+\.\w+\b/g, "")
+		.replace(/\/[\w./_-]{4,}/g, "")
+		.replace(/[A-Z]:\\[\w\\._-]+/g, "")
+		.replace(/[ \t]+/g, " ")
+		.trim();
+}
+
+// Split one chunk off the head of `remaining`. Always ends the chunk on
+// sentence punctuation when possible — F5-TTS hallucinates noise at the tail
+// of unpunctuated fragments since it doesn't know where to stop. If no
+// boundary exists in `limit` chars, split on a space and append a period so
+// F5 has a clean stop signal.
 function splitOne(remaining, limit) {
 	if (remaining.length <= limit) return [remaining, ''];
 	const slice = remaining.substring(0, limit);
-	let splitAt = -1;
-	for (let i = slice.length - 1; i > limit * 0.3; i--) {
+	for (let i = slice.length - 1; i > 0; i--) {
 		const ch = slice[i];
 		if ((ch === '.' || ch === '!' || ch === '?' || ch === '\n') &&
 			(i === slice.length - 1 || slice[i + 1] === ' ' || slice[i + 1] === '\n')) {
-			splitAt = i + 1;
-			break;
+			const splitAt = i + 1;
+			return [remaining.substring(0, splitAt).trim(), remaining.substring(splitAt).trim()];
 		}
 	}
-	if (splitAt === -1) splitAt = slice.lastIndexOf(' ');
+	let splitAt = slice.lastIndexOf(' ');
 	if (splitAt <= 0) splitAt = limit;
-	return [remaining.substring(0, splitAt).trim(), remaining.substring(splitAt).trim()];
+	const head = remaining.substring(0, splitAt).trim();
+	const tail = remaining.substring(splitAt).trim();
+	const punctuatedHead = /[.!?]$/.test(head) ? head : head + ".";
+	return [punctuatedHead, tail];
 }
 
 function splitTextIntoChunks(text) {
-	if (text.length <= TTS_FIRST_CHUNK_MAX) return [text];
+	const cleaned = stripUnspeakable(text);
+	if (cleaned.length <= TTS_FIRST_CHUNK_MAX) return [cleaned];
 
 	const chunks = [];
-	let remaining = text;
+	let remaining = cleaned;
 	let isFirst = true;
 	while (remaining.length > 0) {
 		const limit = isFirst ? TTS_FIRST_CHUNK_MAX : TTS_CHUNK_MAX;
@@ -105,6 +126,8 @@ function splitTextIntoChunks(text) {
 	}
 
 	console.log(`TTS: split ${text.length} chars into ${chunks.length} chunks:`, chunks.map((c, i) => `chunk${i + 1}=${c.length}chars`));
+	// Diagnostic: log full text of each chunk so we can spot garbage being sent to F5.
+	chunks.forEach((c, i) => console.log(`TTS chunk ${i + 1} text:`, JSON.stringify(c)));
 	return chunks;
 }