app.py

import tempfile
import subprocess
import time

from typing import Optional
from AinaTheme import AinaGradioTheme
import gradio as gr
import numpy as np
import torch
import os
from TTS.utils.synthesizer import Synthesizer

from dotenv import load_dotenv

torch.manual_seed(0)
np.random.seed(0)

import json
from copy import deepcopy

import numpy as np
import torch

import torchaudio
import torchaudio.transforms as T

import random

random.seed(0)
torch.manual_seed(0)
np.random.seed(0)

SAMPLE_RATE = 8000

#############################################################################################################

load_dotenv()

MAX_INPUT_TEXT_LEN = int(os.environ.get("MAX_INPUT_TEXT_LEN", default=500))

# Dynamically read model files, exclude 'speakers.pth'
model_files = [f for f in os.listdir(os.getcwd()) if f.endswith('.pth') and f != 'speakers.pth']
# model_files = [f for f in os.listdir(os.path.join(os.getcwd(), 'checkpoints')) if f.endswith('.pth')]
# model_files.sort(key=lambda x: os.path.getmtime(os.path.join(os.getcwd(), x)), reverse=True)

speakers_path = "speakers.pth"
speakers_list = torch.load(speakers_path)
speakers_list = list(speakers_list.keys())
speakers_list = [speaker for speaker in speakers_list]

default_speaker_list = speakers_list  #

# Filtered lists based on dataset
festcat_speakers = [s for s in speakers_list if len(s) == 3]  #
google_speakers = [s for s in speakers_list if 3 < len(s) < 20]  #
commonvoice_speakers = [s for s in speakers_list if len(s) > 20]  #

DEFAULT_SPEAKER_ID = os.environ.get("DEFAULT_SPEAKER_ID", default="pau")
DEFAULT_CHECKPOINT = os.environ.get("DEFAULT_CHECKPOINT", default=model_files[0])
# model_file = model_files[0]  # change this!!

# model_path = os.path.join(os.getcwd(), model_file)
# config_path = os.path.join(os.getcwd(), "config.json")

vocoder_path = None
vocoder_config_path = None

# synthesizer = Synthesizer(
#     model_path, config_path, speakers_path, None, vocoder_path, vocoder_config_path,
# )


def get_phonetic_transcription(text: str):
    try:
        result = subprocess.run(
            ['espeak-ng', '--ipa', '-v', 'ca', text],
            stdout=subprocess.PIPE,
            stderr=subprocess.PIPE,
            text=True,
            check=True
        )
        return result.stdout.strip()
    except subprocess.CalledProcessError as e:
        print(f"An error occurred: {e}")
        return None


def tts_inference(text: str, speaker_idx: str = None, model_file: str=None):

    model_path = os.path.join(os.getcwd(), model_file)
    speakers_file_path = "speakers.pth"
    config_path = "config.json"
    vocoder_path = None
    vocoder_config_path = None

    synthesizer = Synthesizer(model_path, config_path, speakers_path, None, 
                              vocoder_path, vocoder_config_path)
    # synthesize
    if synthesizer is None:
        raise NameError("model not found")
    t1 = time.time()
    wavs = synthesizer.tts(text, speaker_idx)
    # print(type(wavs))
    wavs_den = wavs

    # return output
    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as fp:
        # wavs must be a list of integers
        synthesizer.save_wav(wavs_den, fp)
        t2 = time.time() - t1
        print(round(t2, 2))
        output_audio = fp.name

    return output_audio


title = "🗣️ Catalan Multispeaker TTS Tester 🗣️"
description = """
1️⃣ Enter the text to synthesize.
2️⃣ Select a voice from the dropdown menu.
3️⃣ Enjoy!
"""


def submit_input(input_, speaker_id, model_chkpt):
    output_audio = None
    output_phonetic = None
    if input_ is not None and len(input_) < MAX_INPUT_TEXT_LEN:
        output_audio = tts_inference(input_, speaker_id, model_chkpt)
        output_phonetic = get_phonetic_transcription(input_)
    else:
        gr.Warning(f"Your text exceeds the {MAX_INPUT_TEXT_LEN}-character limit.")
    return output_audio, output_phonetic


def change_interactive(text):
    input_state = text
    if input_state.strip() != "":
        return gr.update(interactive=True)
    else:
        return gr.update(interactive=False)


def clean():
    return (
        None,
        None,
    )


with gr.Blocks(**AinaGradioTheme().get_kwargs()) as app:
    gr.Markdown(f"<h1 style='text-align: center; margin-bottom: 1rem'>{title}</h1>")
    gr.Markdown(description)

    with gr.Row(equal_height=False):

        with gr.Column(variant='panel'):
            input_ = gr.Textbox(
                label="Text",
                value="Introdueix el text a sintetitzar.",
                lines=4
            )

            dataset = gr.Radio(["All", "Festcat", "Google TTS", "CommonVoice"], label="Speakers Dataset",
                               value="All")


            def update_speaker_list(dataset):
                print("Updating speaker list based on dataset:", dataset)
                if dataset == "Festcat":
                    current_speakers = festcat_speakers
                elif dataset == "Google TTS":
                    current_speakers = google_speakers
                elif dataset == "CommonVoice":
                    current_speakers = commonvoice_speakers
                else:
                    current_speakers = speakers_list

                return gr.update(choices=current_speakers, value=current_speakers[0])


            speaker_id = gr.Dropdown(label="Select a voice", choices=speakers_list, value=DEFAULT_SPEAKER_ID,
                                     interactive=True)
            dataset.change(fn=update_speaker_list, inputs=dataset, outputs=speaker_id)

            model_chkpt = gr.Dropdown(label="Select a checkpoint", choices=model_files, value=DEFAULT_CHECKPOINT, interactive=True)

            # model = gr.Dropdown(label="Select a model", choices=model_files, value=DEFAULT_MODEL_FILE_NAME)
            with gr.Row():
                clear_btn = gr.ClearButton(value='Clean', components=[input_])
                # clear_btn = gr.Button(
                #     "Clean",
                # )
                submit_btn = gr.Button(
                    "Submit",
                    variant="primary",
                )
                # use_denoise = gr.Radio(choices=[("Yes", 0), ("No", 1)], value=0)
        with gr.Column(variant='panel'):
            output_audio = gr.Audio(label="Output", type="filepath", autoplay=True, show_share_button=False)
            # output_audio_den = gr.Audio(label="Output denoised", type="filepath", autoplay=False, show_share_button=False)

            output_phonetic = gr.Textbox(label="Phonetic Transcription", readonly=True)

    for button in [submit_btn]:  # clear_btn
        input_.change(fn=change_interactive, inputs=[input_], outputs=button)

    # clear_btn.click(fn=clean, inputs=[], outputs=[input_, output_audio, output_phonetic], queue=False)
    submit_btn.click(fn=submit_input, inputs=[input_, speaker_id, model_chkpt], outputs=[output_audio, output_phonetic])

app.queue(concurrency_count=1, api_open=False)
app.launch(show_api=False, server_name="0.0.0.0", server_port=7860)