feat: raw phonemes

src/cpp/piper.cpp

@@ -69,6 +69,8 @@ void parsePhonemizeConfig(json &configRoot, PhonemizeConfig &phonemizeConfig) {
     auto phonemeTypeStr = configRoot["phoneme_type"].get<std::string>();
     if (phonemeTypeStr == "text") {
       phonemizeConfig.phonemeType = TextPhonemes;
+    } else if (phonemeTypeStr == "raw") {
+      phonemizeConfig.phonemeType = RawPhonemes;
     }
   }
 
@@ -467,7 +469,16 @@ void textToAudio(PiperConfig &config, Voice &voice, std::string text,
   spdlog::debug("Phonemizing text: {}", text);
   std::vector<std::vector<Phoneme>> phonemes;
 
-  if (voice.phonemizeConfig.phonemeType == eSpeakPhonemes) {
+  if (voice.phonemizeConfig.phonemeType == RawPhonemes) {
+    // Use text as raw utf-8 phonemes
+    phonemes.emplace_back();
+    // UTF-8 wide
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    std::wstring wide_text = converter.from_bytes(text);
+    for (wchar_t c : wide_text) {
+      phonemes.back().push_back(c);  // Adds each wide character as a phoneme
+    }
+  } else if (voice.phonemizeConfig.phonemeType == eSpeakPhonemes) {
     // Use espeak-ng for phonemization
     eSpeakPhonemeConfig eSpeakConfig;
     eSpeakConfig.voice = voice.phonemizeConfig.eSpeak.voice;

src/cpp/piper.hpp

@@ -34,7 +34,7 @@ struct PiperConfig {
   std::unique_ptr<tashkeel::State> tashkeelState;
 };
 
-enum PhonemeType { eSpeakPhonemes, TextPhonemes };
+enum PhonemeType { eSpeakPhonemes, TextPhonemes, RawPhonemes };
 
 struct PhonemizeConfig {
   PhonemeType phonemeType = eSpeakPhonemes;

src/python/piper_train/preprocess.py

@@ -103,6 +103,9 @@ def main() -> None:
     parser.add_argument(
         "--debug", action="store_true", help="Print DEBUG messages to the console"
     )
+    parser.add_argument(
+        "--raw-phonemes", action="store_true", help="Raw espeak compatible phonemes"
+    )
     args = parser.parse_args()
 
     if args.single_speaker and (args.speaker_id is not None):
@@ -299,7 +302,10 @@ def phonemize_batch_espeak(
                         utt.text = tashkeel_run(utt.text)
 
                     _LOGGER.debug(utt)
-                    all_phonemes = phonemize_espeak(casing(utt.text), args.language)
+                    if args.raw_phonemes:
+                        all_phonemes = [utt.text.split()]
+                    else:
+                        all_phonemes = phonemize_espeak(casing(utt.text), args.language)
 
                     # Flatten
                     utt.phonemes = [