app.py

import gradio as gr
import tempfile
from TTS.api import TTS
from huggingface_hub import hf_hub_download
import torch

CUDA = torch.cuda.is_available()

REPO_ID = "collectivat/catotron-ona"

VOICE_CONVERSION_MODELS = {
    'freevc24': 'voice_conversion_models/multilingual/vctk/freevc24',
    'openvoice_v1': 'voice_conversion_models/multilingual/multi-dataset/openvoice_v1',
    'openvoice_v2': 'voice_conversion_models/multilingual/multi-dataset/openvoice_v2',
}

my_title = "Catotron Text-to-Speech with Voice Conversion"
my_description = "This space allows speaker conversion on Fast Speech based 🐸 [Catotron](https://huggingface.co/collectivat/catotron-ona)." 

my_examples = [
    ["Catotron, síntesi de la parla obert i lliure en català.", True, None, 'freevc24'],
    ["Leonor Ferrer Girabau va ser una delineant, mestra i activista barcelonina, nascuda al carrer actual de la Concòrdia del Poble-sec, que es va convertir en la primera dona a obtenir el títol de delineant a Catalunya i a l'estat.", True, None, 'freevc24'],
    ["S'espera un dia anticiclònic amb temperatures suaus i vent fluix.", False, None, 'freevc24']
]

my_inputs = [
    gr.Textbox(lines=5, label="Input Text"),
    gr.Checkbox(label="Split Sentences", value=False),
    gr.Audio(type="filepath", label="Speaker audio for voice cloning (optional)"),
    gr.Dropdown(label="Voice Conversion Model", choices=list(VOICE_CONVERSION_MODELS.keys())),
]

my_outputs = gr.Audio(type="filepath", label="Output Audio", autoplay=True)

best_model_path = hf_hub_download(repo_id=REPO_ID, filename="fast-speech_best_model.pth") 
config_path = hf_hub_download(repo_id=REPO_ID, filename="fast-speech_config.json")
vocoder_model = hf_hub_download(repo_id=REPO_ID, filename="ljspeech--hifigan_v2_model_file.pth")
vocoder_config = hf_hub_download(repo_id=REPO_ID, filename="ljspeech--hifigan_v2_config.json")

api = TTS(model_path=best_model_path, config_path=config_path, vocoder_path=vocoder_model, vocoder_config_path=vocoder_config).to("cuda" if CUDA else "cpu")

# pre-download voice conversion models
for model in VOICE_CONVERSION_MODELS.values():
    api.load_vc_model_by_name(model, gpu=CUDA)

def tts(text: str, split_sentences: bool = False, speaker_wav: str = None, voice_cv_model: str = 'freevc24'):
    # replace oov characters
    text = text.replace("\n", ". ")
    text = text.replace("(", ",")
    text = text.replace(")", ",")
    text = text.replace(";", ",")

    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as fp:
        if speaker_wav:
            api.load_vc_model_by_name(VOICE_CONVERSION_MODELS[voice_cv_model], gpu=CUDA)
            api.tts_with_vc_to_file(text, speaker_wav=speaker_wav, file_path=fp.name, split_sentences=split_sentences)
        else:
            api.tts_to_file(text, file_path=fp.name, split_sentences=split_sentences)

    return fp.name

iface = gr.Interface(
    fn=tts, 
    inputs=my_inputs, 
    outputs=my_outputs, 
    title=my_title, 
    description=my_description, 
    examples=my_examples,
    cache_examples=True
)
iface.launch()