StyleTTSEngine language in phonemizer is hardcoded to English (en-us)

[kohai-channel]

Hi, I noticed that in StyleTTSEngine, the language used in the phonemizer is hardcoded to English (en-us), which prevents using the engine with models trained for other languages.

Specifically, this line in the code:

# Initialize phonemizer
self.global_phonemizer = phonemizer.backend.EspeakBackend(language='en-us', 
                                                          preserve_punctuation=True,
                                                          with_stress=True)

I have a fine-tuned model that speaks Spanish, but I can't use it properly with StyleTTSEngine because of this limitation. It would be great if the phonemizer language could be configurable or inferred from the model's settings.

Thanks!

[KoljaB]

I'll take care of this in the next release. Do you by any chance know good resources how to fine tune on another language? Tried for german but failed so far...

[sachin7695]

@KoljaB i am using styletts2 with realtimeTTS engine for my native language Hindi which i fine tuned on my own data !! i am currently facing issue with the language_switch ( hi ) nəmˈʌsteː dˈʊnɪjˌaː ( en-us ) , ( hi ) sʈˌaːɪlʈiʈiˈeːs ˈɔɖɪjˌoː kaː pəɾˈiːkʃəɳ kˈɪjaː ɟˈaː ɾˌəhaː hɛː ( en-us ) i tried to remove it in this way self.global_phonemizer = phonemizer.backend.EspeakBackend( language='hi', preserve_punctuation=True, with_stress=False, language_switch='remove-flags', # This should remove the flags words_mismatch='ignore', # Add these additional settings to suppress warnings punctuation_marks=';:,.!?¡¿—…"«»""', strip=True )
having this language_switch='remove-flags', # This should remove the flags still its synthesizing in this way!! where i am doing wrong??? also i have created styletts2 tts service class supported by RealtimeTTS engine for pipecat !! whole logs

2025-05-29 11:36:16.996 | INFO     | __main__:start:159 - StyleTTSService#0: TextToAudioStream created successfully
2025-05-29 11:36:16.996 | INFO     | __main__:start:166 - StyleTTSService#0: StyleTTS initialization completed successfully in 8.82s
2025-05-29 11:36:16.996 | INFO     | __main__:run_tts:226 - StyleTTSService#0: Starting TTS generation for: [नमस्ते दुनिया, स्टाइलटीटीएस ऑडियो का परीक्षण किया ...]
2025-05-29 11:36:16.996 | DEBUG    | __main__:run_tts:241 - StyleTTSService#0: Processing text: [नमस्ते दुनिया, स्टाइलटीटीएस ऑडियो का परीक्षण किया जा रहा है]
2025-05-29 11:36:16.996 | DEBUG    | __main__:run_tts:249 - StyleTTSService#0: Starting audio streaming...
2025-05-29 11:36:16.996 | DEBUG    | __main__:_stream_audio_realtime:269 - StyleTTSService#0: Setting up 200ms buffered audio streaming for text: [नमस्ते दुनिया, स्टाइलटीटीएस ऑड...]
2025-05-29 11:36:16.996 | INFO     | __main__:run_synthesis:308 - StyleTTSService#0: Starting synthesis thread for text: [नमस्ते दुनिया, स्टाइलटीटीएस ऑड...]
⚡ synthesizing → 'नमस्ते दुनिया, स्टाइलटीटीएस ऑडियो का परीक्षण किया जा रहा है'
WARNING:phonemizer:2 utterances containing language switches on lines 1, 2
WARNING:phonemizer:extra phones may appear in the "en-us" phoneset
WARNING:phonemizer:language switch flags have been kept (applying "keep-flags" policy)
( hi ) nəmˈʌsteː dˈʊnɪjˌaː ( en-us ) , ( hi ) sʈˌaːɪlʈiʈiˈeːs ˈɔɖɪjˌoː kaː pəɾˈiːkʃəɳ kˈɪjaː ɟˈaː ɾˌəhaː hɛː ( en-us )

New Padding length bert_dur_2: 109
SYNTHESIS FINISHED

[KoljaB]

I'm sorry I have no time for anything currently. That can take a while until I can look into that.