app.py

import gradio as gr
import torch
import torchaudio
import tempfile

import numpy as np
from nemo.collections.tts.models import FastPitchModel
from nemo.collections.tts.models import HifiGanModel
from nemo.collections.tts.models import MixerTTSModel

from transformers import pipeline

# spec_generator_2 = MixerTTSModel.from_pretrained("tts_en_lj_mixerttsx")
# model1 = HifiGanModel.from_pretrained(model_name="tts_en_lj_hifigan_ft_mixerttsx")

def greet(name):
    return "Hello " + name + "!!"

def run(spec_generator, voc_model, pipe):

    def generate_tts(text: str, speaker: int = 0):
        sr = 44100
        parsed = spec_generator.parse(text)
        spectrogram = spec_generator.generate_spectrogram(tokens=parsed, speaker=speaker)
        audio = voc_model.convert_spectrogram_to_audio(spec=spectrogram)
        
        return gr.Audio.update(sr, audio.squeeze(0).cpu().numpy())

    demo = gr.Interface(
        fn=generate_tts,
        inputs=[gr.Textbox(value="This is a test.", label="Text to Synthesize"),
                gr.Slider(0, 10, step=1, label="Speaker")],
        outputs=gr.Audio(label="Output", type="numpy"),
        allow_flagging=False,
    )

    demo.launch(server_name="0.0.0.0", server_port=7860)

if __name__ == "__main__":

    spec_generator = FastPitchModel.from_pretrained("tts_en_fastpitch_multispeaker")
    spec_generator.eval()
    voc_model = HifiGanModel.from_pretrained(model_name="tts_en_hifitts_hifigan_ft_fastpitch")
    voc_model.eval()

    pipe = pipeline("text-to-speech", model="suno/bark-small")
    
    run(spec_generator, voc_model, pipe)