Fix spliting text by languages for kokoro tts. (#1849)

sherpa-onnx/c-api/cxx-api.cc

@@ -270,7 +270,8 @@ OfflineStream OfflineRecognizer::CreateStream() const {
   return OfflineStream{s};
 }
 
-OfflineStream OfflineRecognizer::CreateStream(const std::string &hotwords) const {
+OfflineStream OfflineRecognizer::CreateStream(
+    const std::string &hotwords) const {
   auto s = SherpaOnnxCreateOfflineStreamWithHotwords(p_, hotwords.c_str());
   return OfflineStream{s};
 }

sherpa-onnx/csrc/CMakeLists.txt

@@ -549,6 +549,7 @@ if(SHERPA_ONNX_ENABLE_TESTS)
     context-graph-test.cc
     packed-sequence-test.cc
     pad-sequence-test.cc
+    regex-lang-test.cc
     slice-test.cc
     stack-test.cc
     text-utils-test.cc

sherpa-onnx/csrc/kokoro-multi-lang-lexicon.cc

@@ -4,9 +4,7 @@
 
 #include "sherpa-onnx/csrc/kokoro-multi-lang-lexicon.h"
 
-#include <codecvt>
 #include <fstream>
-#include <locale>
 #include <regex>  // NOLINT
 #include <sstream>
 #include <strstream>
@@ -22,6 +20,8 @@
 #include "rawfile/raw_file_manager.h"
 #endif
 
+#include <codecvt>
+
 #include "cppjieba/Jieba.hpp"
 #include "espeak-ng/speak_lib.h"
 #include "phoneme_ids.hpp"
@@ -37,20 +37,6 @@ void CallPhonemizeEspeak(const std::string &text,
                          piper::eSpeakPhonemeConfig &config,  // NOLINT
                          std::vector<std::vector<piper::Phoneme>> *phonemes);
 
-static std::wstring ToWideString(const std::string &s) {
-  // see
-  // https://stackoverflow.com/questions/2573834/c-convert-string-or-char-to-wstring-or-wchar-t
-  std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
-  return converter.from_bytes(s);
-}
-
-static std::string ToString(const std::wstring &s) {
-  // see
-  // https://stackoverflow.com/questions/2573834/c-convert-string-or-char-to-wstring-or-wchar-t
-  std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
-  return converter.to_bytes(s);
-}
-
 class KokoroMultiLangLexicon::Impl {
  public:
   Impl(const std::string &tokens, const std::string &lexicon,
@@ -103,37 +89,42 @@ class KokoroMultiLangLexicon::Impl {
 
     // https://en.cppreference.com/w/cpp/regex
     // https://stackoverflow.com/questions/37989081/how-to-use-unicode-range-in-c-regex
-    std::string expr =
-        "([;:,.?!'\"…\\(\\)“”])|([\\u4e00-\\u9fff]+)|([äöüßÄÖÜ\\u0000-\\u007f]+"
-        ")";
+    std::string expr_chinese = "([\\u4e00-\\u9fff]+)";
+    std::string expr_not_chinese = "([^\\u4e00-\\u9fff]+)";
+
+    std::string expr_both = expr_chinese + "|" + expr_not_chinese;
 
     auto ws = ToWideString(text);
-    std::wstring wexpr = ToWideString(expr);
-    std::wregex we(wexpr);
+    std::wstring wexpr_both = ToWideString(expr_both);
+    std::wregex we_both(wexpr_both);
+
+    std::wstring wexpr_zh = ToWideString(expr_chinese);
+    std::wregex we_zh(wexpr_zh);
 
-    auto begin = std::wsregex_iterator(ws.begin(), ws.end(), we);
+    auto begin = std::wsregex_iterator(ws.begin(), ws.end(), we_both);
     auto end = std::wsregex_iterator();
 
     std::vector<TokenIDs> ans;
 
     for (std::wsregex_iterator i = begin; i != end; ++i) {
       std::wsmatch match = *i;
       std::wstring match_str = match.str();
+
       auto ms = ToString(match_str);
       uint8_t c = reinterpret_cast<const uint8_t *>(ms.data())[0];
 
       std::vector<std::vector<int32_t>> ids_vec;
-
-      if (c < 0x80) {
+      if (std::regex_match(match_str, we_zh)) {
         if (debug_) {
-          SHERPA_ONNX_LOGE("Non-Chinese: %s", ms.c_str());
+          SHERPA_ONNX_LOGE("Chinese: %s", ms.c_str());
         }
-        ids_vec = ConvertEnglishToTokenIDs(ms, meta_data_.voice);
+        ids_vec = ConvertChineseToTokenIDs(ms);
       } else {
         if (debug_) {
-          SHERPA_ONNX_LOGE("Chinese: %s", ms.c_str());
+          SHERPA_ONNX_LOGE("Non-Chinese: %s", ms.c_str());
         }
-        ids_vec = ConvertChineseToTokenIDs(ms);
+
+        ids_vec = ConvertEnglishToTokenIDs(ms, meta_data_.voice);
       }
 
       for (const auto &ids : ids_vec) {
@@ -315,9 +306,10 @@ class KokoroMultiLangLexicon::Impl {
         this_sentence.push_back(space_id);
       } else {
         if (debug_) {
-          SHERPA_ONNX_LOGE("Use espeak-ng to handle the OOV: '%s'", word.c_str());
+          SHERPA_ONNX_LOGE("Use espeak-ng to handle the OOV: '%s'",
+                           word.c_str());
         }
-        
+
         piper::eSpeakPhonemeConfig config;
 
         config.voice = voice;

sherpa-onnx/csrc/regex-lang-test.cc

@@ -0,0 +1,86 @@
+// sherpa-onnx/csrc/regex-lang-test.cc
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#include <regex>  // NOLINT
+
+#include "gtest/gtest.h"
+#include "sherpa-onnx/csrc/text-utils.cc"
+
+namespace sherpa_onnx {
+
+static void TestLang(const std::string &expr, const std::string &text,
+                     const std::vector<std::string> &expected) {
+  auto ws = ToWideString(text);
+  std::wstring wexpr = ToWideString(expr);
+  std::wregex we(wexpr);
+
+  auto begin = std::wsregex_iterator(ws.begin(), ws.end(), we);
+  auto end = std::wsregex_iterator();
+  int32_t k = 0;
+  for (std::wsregex_iterator i = begin; i != end; ++i) {
+    std::wsmatch match = *i;
+    std::wstring match_str = match.str();
+    auto ms = ToString(match_str);
+    std::cout << ms << "\n";
+    EXPECT_EQ(ms, expected[k]);
+    k++;
+  }
+  EXPECT_EQ(k, expected.size());
+}
+
+TEST(German, Case1) {
+  std::cout << "----------Test German----------";
+  // see https://character-table.netlify.app/german/
+  std::string expr =
+      "([\\u0020-\\u005f\\u0061-"
+      "\\u007d\\u00a0\\u00a7\\u00a9\\u00ab\\u00bb\\u00c4\\u00d6\\u00dc\\u00df\\"
+      "u00e4\\u00f6\\u00fc\\u2010-\\u2011\\u2013-"
+      "\\u2014\\u2018\\u201a\\u201c\\u201e\\u2026\\u2030\\u20ac]+)";
+
+  std::string text =
+      "开始Übeltäter übergibt Ärzten 中间öfters äußerst ätzende Öle结束3€";
+
+  std::vector<std::string> expected = {"Übeltäter übergibt Ärzten ",
+                                       "öfters äußerst ätzende Öle", "3€"};
+
+  TestLang(expr, text, expected);
+}
+
+TEST(French, Case1) {
+  std::string expr =
+      "([\\u0020-\\u005f\\u0061-"
+      "\\u007a\\u007c\\u00a0\\u00a7\\u00a9\\u00ab\\u00b2-"
+      "\\u00b3\\u00bb\\u00c0\\u00c2\\u00c6-\\u00cb\\u00ce-"
+      "\\u00cf\\u00d4\\u00d9\\u00db-\\u00dc\\u00e0\\u00e2\\u00e6-"
+      "\\u00eb\\u00ee-\\u00ef\\u00f4\\u00f9\\u00fb-\\u00fc\\u00ff\\u0152-"
+      "\\u0153\\u0178\\u02b3\\u02e2\\u1d48-\\u1d49\\u2010-\\u2011\\u2013-"
+      "\\u2014\\u2019\\u201c-\\u201d\\u2020-\\u2021\\u2026\\u202f-"
+      "\\u2030\\u20ac\\u2212]+)";
+  std::string text =
+      "L'été, 一avec son ciel bleuâtre, 二est un moment où, 三Noël, maçon";
+  std::vector<std::string> expected = {
+      "L'été, ",
+      "avec son ciel bleuâtre, ",
+      "est un moment où, ",
+      "Noël, maçon",
+  };
+  TestLang(expr, text, expected);
+}
+
+TEST(English, Case1) {
+  // https://character-table.netlify.app/english/
+  std::string expr =
+      "([\\u0020-\\u005f\\u0061-\\u007a\\u007c\\u00a0\\u00a7\\u00a9\\u2010-"
+      "\\u2011\\u2013-\\u2014\\u2018-\\u2019\\u201c-\\u201d\\u2020-"
+      "\\u2021\\u2026\\u2030\\u2032-\\u2033\\u20ac]+)";
+  std::string text = "一how are you doing? 二Thank you!";
+
+  std::vector<std::string> expected = {
+      "how are you doing? ",
+      "Thank you!",
+  };
+  TestLang(expr, text, expected);
+}
+
+}  // namespace sherpa_onnx

sherpa-onnx/csrc/text-utils-test.cc

@@ -8,6 +8,14 @@
 
 namespace sherpa_onnx {
 
+TEST(ToLowerCase, WideString) {
+  std::string text =
+      "Hallo! Übeltäter übergibt Ärzten öfters äußerst ätzende Öle 3€";
+  auto t = ToLowerCase(text);
+  std::cout << text << "\n";
+  std::cout << t << "\n";
+}
+
 TEST(RemoveInvalidUtf8Sequences, Case1) {
   std::vector<uint8_t> v = {
       0xe4, 0xbb, 0x8a,                                  // 今

sherpa-onnx/csrc/text-utils.cc

@@ -8,8 +8,11 @@
 #include <algorithm>
 #include <cassert>
 #include <cctype>
+#include <codecvt>
 #include <cstdint>
+#include <cwctype>
 #include <limits>
+#include <locale>
 #include <sstream>
 #include <string>
 #include <unordered_map>
@@ -389,17 +392,74 @@ std::vector<std::string> SplitUtf8(const std::string &text) {
 }
 
 std::string ToLowerCase(const std::string &s) {
-  std::string ans(s.size(), 0);
-  std::transform(s.begin(), s.end(), ans.begin(),
-                 [](unsigned char c) { return std::tolower(c); });
-  return ans;
+  return ToString(ToLowerCase(ToWideString(s)));
 }
 
 void ToLowerCase(std::string *in_out) {
   std::transform(in_out->begin(), in_out->end(), in_out->begin(),
                  [](unsigned char c) { return std::tolower(c); });
 }
 
+std::wstring ToLowerCase(const std::wstring &s) {
+  std::wstring ans(s.size(), 0);
+  std::transform(s.begin(), s.end(), ans.begin(), [](wchar_t c) -> wchar_t {
+    switch (c) {
+      // French
+      case L'À':
+        return L'à';
+      case L'Â':
+        return L'â';
+      case L'Æ':
+        return L'æ';
+      case L'Ç':
+        return L'ç';
+      case L'È':
+        return L'è';
+      case L'É':
+        return L'é';
+      case L'Ë':
+        return L'ë';
+      case L'Î':
+        return L'î';
+      case L'Ï':
+        return L'ï';
+      case L'Ô':
+        return L'ô';
+      case L'Ù':
+        return L'ù';
+      case L'Û':
+        return L'û';
+      case L'Ü':
+        return L'ü';
+
+      // others
+      case L'Á':
+        return L'á';
+      case L'Í':
+        return L'í';
+      case L'Ó':
+        return L'ó';
+      case L'Ú':
+        return L'ú';
+      case L'Ñ':
+        return L'ñ';
+      case L'Ì':
+        return L'ì';
+      case L'Ò':
+        return L'ò';
+      case L'Ä':
+        return L'ä';
+      case L'Ö':
+        return L'ö';
+        // TODO(fangjun): Add more
+
+      default:
+        return std::towlower(c);
+    }
+  });
+  return ans;
+}
+
 static inline bool InRange(uint8_t x, uint8_t low, uint8_t high) {
   return low <= x && x <= high;
 }
@@ -625,4 +685,18 @@ std::string Gb2312ToUtf8(const std::string &text) {
 }
 #endif
 
+std::wstring ToWideString(const std::string &s) {
+  // see
+  // https://stackoverflow.com/questions/2573834/c-convert-string-or-char-to-wstring-or-wchar-t
+  std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+  return converter.from_bytes(s);
+}
+
+std::string ToString(const std::wstring &s) {
+  // see
+  // https://stackoverflow.com/questions/2573834/c-convert-string-or-char-to-wstring-or-wchar-t
+  std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+  return converter.to_bytes(s);
+}
+
 }  // namespace sherpa_onnx

sherpa-onnx/csrc/text-utils.h

@@ -124,6 +124,8 @@ std::vector<std::string> SplitUtf8(const std::string &text);
 std::string ToLowerCase(const std::string &s);
 void ToLowerCase(std::string *in_out);
 
+std::wstring ToLowerCase(const std::wstring &s);
+
 std::string RemoveInvalidUtf8Sequences(const std::string &text,
                                        bool show_debug_msg = false);
 
@@ -139,6 +141,10 @@ bool IsGB2312(const std::string &text);
 std::string Gb2312ToUtf8(const std::string &text);
 #endif
 
+std::wstring ToWideString(const std::string &s);
+
+std::string ToString(const std::wstring &s);
+
 }  // namespace sherpa_onnx
 
 #endif  // SHERPA_ONNX_CSRC_TEXT_UTILS_H_