From 3be42af9bbb5b239dff08b24e52f635dce02cbca Mon Sep 17 00:00:00 2001
From: enne2 <me@enne2.net>
Date: Thu, 13 Aug 2026 19:54:50 +0200
Subject: [PATCH] tts-server: normalizza i numeri interi 5+ cifre in parole
 italiane (Qwen3-TTS li leggeva cifra per cifra)

---
 src/text-normalize.h | 123 +++++++++++++++++++++++++++++++++++++++++++
 tools/tts-server.cpp |   5 +-
 2 files changed, 127 insertions(+), 1 deletion(-)
 create mode 100644 src/text-normalize.h

diff --git a/src/text-normalize.h b/src/text-normalize.h
new file mode 100644
index 0000000..13b6e65
--- /dev/null
+++ b/src/text-normalize.h
@@ -0,0 +1,123 @@
+// text-normalize.h: normalizzazione testo per TTS.
+// Qwen3-TTS legge i numeri interi di 5+ cifre cifra per cifra
+// ("12345" -> "un, due, tre, quattro, cinque"). Questa funzione li
+// converte in parole italiane prima della sintesi.
+#pragma once
+
+#include <cstdint>
+#include <string>
+
+namespace text_normalize {
+
+static bool is_alnum_utf8(char c) {
+    unsigned char u = (unsigned char) c;
+    return (u >= '0' && u <= '9') || (u >= 'A' && u <= 'Z') || (u >= 'a' && u <= 'z') || u >= 0x80;
+}
+
+// Converte un intero (0..999.999.999.999) in parole italiane.
+static std::string number_to_words(uint64_t n) {
+    if (n == 0) return "zero";
+
+    static const char * units[] = {"", "uno", "due", "tre", "quattro", "cinque", "sei", "sette", "otto", "nove"};
+    static const char * teens[] = {"dieci", "undici", "dodici", "tredici", "quattordici", "quindici", "sedici",
+                                   "diciassette", "diciotto", "diciannove"};
+    static const char * tens[]  = {"", "", "venti", "trenta", "quaranta", "cinquanta", "sessanta", "settanta",
+                                   "ottanta", "novanta"};
+
+    // Converte un gruppo di 3 cifre (0-999).
+    auto group3 = [&](int g) -> std::string {
+        if (g == 0) return "";
+        std::string out;
+        int h = g / 100, rest = g % 100;
+        if (h > 0) {
+            out += (h == 1) ? "cento" : (std::string(units[h]) + "cento");
+        }
+        if (rest >= 10 && rest < 20) {
+            out += teens[rest - 10];
+        } else {
+            int t = rest / 10, u = rest % 10;
+            if (t > 1) {
+                std::string dec = tens[t];
+                if (u == 1 || u == 8) {
+                    dec.pop_back(); // venti+uno=ventuno, venti+otto=ventotto
+                    out += dec + units[u];
+                } else if (u == 3) {
+                    out += dec + "tr\xc3\xa9"; // ventitré (accento)
+                } else if (u > 0) {
+                    out += dec + units[u];
+                } else {
+                    out += dec;
+                }
+            } else if (t == 1) {
+                out += teens[u];
+            } else if (u > 0) {
+                out += units[u];
+            }
+        }
+        return out;
+    };
+
+    static const char * scale_singular[] = {"", "mille", "un milione", "un miliardo"};
+    static const char * scale_plural[]   = {"", "mila", "milioni", "miliardi"};
+
+    int groups[4] = {0, 0, 0, 0};
+    uint64_t v = n;
+    for (int i = 0; i < 4 && v > 0; i++) {
+        groups[i] = (int) (v % 1000);
+        v /= 1000;
+    }
+
+    std::string out;
+    bool first = true;
+    for (int i = 3; i >= 0; i--) {
+        int g = groups[i];
+        if (g == 0) continue;
+        std::string words = group3(g);
+        if (i > 0) {
+            if (g == 1) {
+                words = scale_singular[i];
+            } else {
+                words += " " + std::string(scale_plural[i]);
+            }
+        }
+        if (!first) out += " ";
+        out += words;
+        first = false;
+    }
+    return out;
+}
+
+// Normalizza i numeri interi di 5+ cifre in parole italiane.
+// Non tocca: date, orari, decimali, codici, numeri < 5 cifre, numeri > 12 cifre.
+static void normalize_numbers(std::string & s) {
+    std::string out;
+    out.reserve(s.size() + 32);
+    size_t i = 0;
+    const size_t n = s.size();
+    while (i < n) {
+        unsigned char c = (unsigned char) s[i];
+        if (c >= '0' && c <= '9') {
+            size_t j = i;
+            while (j < n && s[j] >= '0' && s[j] <= '9') j++;
+            size_t len = j - i;
+            bool left_ok  = (i == 0) || !is_alnum_utf8(s[i - 1]);
+            bool right_ok = (j == n) || !is_alnum_utf8(s[j]);
+            if (len >= 5 && len <= 12 && left_ok && right_ok) {
+                uint64_t val = 0;
+                for (size_t k = i; k < j; k++) {
+                    val = val * 10 + (uint64_t) (s[k] - '0');
+                }
+                out += number_to_words(val);
+            } else {
+                out.append(s, i, len);
+            }
+            i = j;
+        } else {
+            out += s[i];
+            i++;
+        }
+    }
+    s.swap(out);
+}
+
+} // namespace text_normalize
diff --git a/tools/tts-server.cpp b/tools/tts-server.cpp
index 2e0ccbd..4bdb1f4 100644
--- a/tools/tts-server.cpp
+++ b/tools/tts-server.cpp
@@ -7,6 +7,7 @@
 
 #include "qwen.h"
 #include "rvq-file.h"
+#include "text-normalize.h"
 #include "version.h"
 
 #include <cmath>
@@ -222,7 +223,9 @@ int main(int argc, char ** argv) {
     be.synthesize = [q, &lang](const tts_request & req, const tts_sink & sink, std::string & err) -> int {
         struct qt_tts_params p;
         qt_tts_default_params(&p);
-        p.text = req.input.c_str();
+        std::string normalized = req.input;
+        text_normalize::normalize_numbers(normalized);
+        p.text = normalized.c_str();
         p.lang = lang.c_str();
 
         // Copy the registered voice latents out under the lock: the