Yourtts: AssertionError: [!] There are duplicate characters in the character set. {' ', ')', ';', '.', "'", '!', ':', '-', '(', ',', '?'}

[suyi00516]

I'm encountering an AssertionError: [!] There are duplicate characters in the character set. when trying to load a YourTTS model (specifically from tts_models--multilingual--multi-dataset--your_tts) locally using TTS.config.load_config followed by TTS.tts.models.setup_model. The error originates in the TTS.tts.utils.text.characters.Graphemes class during vocabulary creation.

CoquiTTS version:0.22.0

The config.json (which I believe is the official one downloaded with the model) contains the following relevant sections:
{
// ... other config ...
"text_cleaner": "multilingual_cleaners",
"use_phonemes": false,
"characters": {
"pad": "_",
"eos": "&",
"bos": "*",
"characters": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz...relevant_symbols...!'(),-.:;? ", // Ends with punctuation
"punctuations": "!'(),-.:;? ",
"phonemes": "...", // phoneme list
"unique": true
},
"model_args": {
"num_chars": 165,
// ... other model_args ...
"use_speaker_encoder_as_loss": true,
"d_vector_file": "/path/to/original/speakers.json_in_config",
// ...
}
}
The characters string in the config unfortunately includes a tail of punctuation marks that are also listed in the punctuations string, leading to the "duplicate characters" error: {' ', ')', ';', '.', "'", '!', ':', '-', '(', ',', '?'}. The characters.unique flag is set to true.
Problem:
When setup_model(C) is called, it eventually initializes TTSTokenizer and Graphemes. The Graphemes class's _create_vocab method throws the assertion error due to these duplicate characters.
If I manually edit config.json to remove the overlapping punctuation from the characters.characters string (leaving them only in characters.punctuations), the AssertionError is resolved. However, this then causes a size mismatch for text_encoder.emb.weight error during model weight loading, because the vocabulary size derived from the modified (shorter) character/punctuation lists no longer matches model_args.num_chars (which is 165). This suggests that num_chars: 165 was defined based on a vocabulary that either included these duplicates or handled their deduplication in a specific way that is not happening in my current setup.
I am trying to load the model locally to use the model.voice_conversion() method with externally computed speaker embeddings, similar to how it's done in some official YourTTS Colab demos. The demos, when using what seems to be the same config.json, do not run into this character duplication error during model setup. The demo code also modifies the loaded config C slightly before calling setup_model(C) (e.g., C.model_args['use_speaker_encoder_as_loss'] = False). I have tried to replicate this modification in my local script, but the character duplication error persists.

[eginhard]

Can you try this with our fork, available via pip install coqui-tts? This repo is not maintained anymore. The following works fine for me there in version 0.26.0:

from TTS.api import TTS
tts = TTS("tts_models/multilingual/multi-dataset/your_tts")
_ = tts.tts("this is a test.", speaker="female-en-5", language="en")

If it still doesn't work, can you share the exact code you're running and the full error message?

[suyi00516]

Thank you for your suggestion to try the coqui-tts fork and the TTS.api.
I'd like to clarify my current setup and the specific challenge I'm encountering, which might be slightly different from a standard TTS.api usage for text-to-speech.
I am working with the tts_models/multilingual/multi-dataset/your_tts model, using locally downloaded model files (weights and config.json). My goal is to utilize the model.voice_conversion() method with externally computed speaker embeddings. For this, I've been loading the model using TTS.config.load_config followed by TTS.tts.models.setup_model (from Coqui-TTS version 0.22.0).
The issue arises from the official config.json provided with this model（mentioned before） I manually edit the config.json to remove the overlapping punctuation from the characters.characters string (leaving them only in characters.punctuations), the "duplicate characters" AssertionError is resolved.
However, this modification then leads to a new error during model weight loading: a size mismatch for text_encoder.emb.weight. The model_args.num_chars (which is 165 in the config) no longer matches the vocabulary size derived from the modified (shorter) character list. This strongly suggests the pretrained text_encoder.emb.weight was created based on a vocabulary that included these duplicates or handled their deduplication in a way that setup_model with the modified config doesn't replicate.

my running code:
# Necessary imports from Coqui TTS
from TTS.config import load_config
from TTS.tts.models import setup_model
# from TTS.utils.audio import AudioProcessor # Not strictly needed to trigger this specific character error

  # --- Configuration ---
  # Path to the local YourTTS config.json file
  # This config.json is assumed to have the problematic character definitions
  # as described in the issue (duplicates between 'characters.characters' and 'characters.punctuations')
  YOURTTS_CONFIG_PATH = "path/to/your/local/yourtts_model/config.json" 
  
  # --- Code to trigger the error ---
  print(f"Loading YourTTS config from: {YOURTTS_CONFIG_PATH}")
  try:
      C = load_config(YOURTTS_CONFIG_PATH)
  
      # The following line is where the error typically occurs during
      # TTSTokenizer and Graphemes class initialization due to character set issues.
      print("Attempting to setup model...")
      model = setup_model(C) 
      print("Model setup successful (unexpected with problematic config).")
  
  except AssertionError as e:
      print(f"AssertionError during model setup: {e}")
      print("This error is typically due to duplicate characters in the character set defined in config.json.")
  except Exception as e:
      print(f"An unexpected error occurred: {e}")

[eginhard]

Ok, I think you get this error because you're using an older version of the model and config. YourTTS shouldn't actually trigger this duplicate check at all. You can use the API to automatically download the correct model or find the up-to-date URLs here: https://github.com/idiap/coqui-ai-TTS/blob/97e4deb58c86dfa222e90af36cf9f2d88dfbcf17/TTS/.models.json#L38

Note that you can still use the API to handle the model loading for you, you can then access the underlying model like this:

from TTS.api import TTS
tts = TTS("tts_models/multilingual/multi-dataset/your_tts")
# or: tts = TTS(model_path="...", config_path="...")
model = tts.synthesizer.tts_model

[suyi00516]

Thank you so much for help, the problem is solved.