app.py

import spaces
from kokoro import KModel, KPipeline
import gradio as gr
import os
import random
import torch

IS_DUPLICATE = not os.getenv('SPACE_ID', '').startswith('igortamara/')
CUDA_AVAILABLE = torch.cuda.is_available()
if not IS_DUPLICATE:
    import kokoro
    import misaki
    print('DEBUG', kokoro.__version__, CUDA_AVAILABLE, misaki.__version__)

CHAR_LIMIT = None if IS_DUPLICATE else 5000
models = {gpu: KModel().to('cuda' if gpu else 'cpu').eval() for gpu in [False] + ([True] if CUDA_AVAILABLE else [])}
pipelines = {lang_code: KPipeline(lang_code=lang_code, model=False) for lang_code in 'e'}

@spaces.GPU(duration=30)
def forward_gpu(ps, ref_s, speed):
    return models[True](ps, ref_s, speed)

def generate_first(text, voice='ef_dora', speed=1, use_gpu=CUDA_AVAILABLE):
    text = text if CHAR_LIMIT is None else text.strip()[:CHAR_LIMIT]
    pipeline = pipelines[voice[0]]
    pack = pipeline.load_voice(voice)
    use_gpu = use_gpu and CUDA_AVAILABLE
    for _, ps, _ in pipeline(text, voice, speed):
        ref_s = pack[len(ps)-1]
        try:
            if use_gpu:
                audio = forward_gpu(ps, ref_s, speed)
            else:
                audio = models[False](ps, ref_s, speed)
        except gr.exceptions.Error as e:
            if use_gpu:
                gr.Warning(str(e))
                gr.Info('Intentando con CPU. Para evitar este error, cambie el Hardware a CPU.')
                audio = models[False](ps, ref_s, speed)
            else:
                raise gr.Error(e)
        return (24000, audio.numpy()), ps
    return None, ''

# Arena API
def predict(text, voice='ef_dora', speed=1):
    return generate_first(text, voice, speed, use_gpu=False)[0]

def tokenize_first(text, voice='ef_dora'):
    pipeline = pipelines[voice[0]]
    for _, ps, _ in pipeline(text, voice):
        return ps
    return ''

def generate_all(text, voice='ef_dora', speed=1, use_gpu=CUDA_AVAILABLE):
    text = text if CHAR_LIMIT is None else text.strip()[:CHAR_LIMIT]
    pipeline = pipelines[voice[0]]
    pack = pipeline.load_voice(voice)
    use_gpu = use_gpu and CUDA_AVAILABLE
    first = True
    for _, ps, _ in pipeline(text, voice, speed):
        ref_s = pack[len(ps)-1]
        try:
            if use_gpu:
                audio = forward_gpu(ps, ref_s, speed)
            else:
                audio = models[False](ps, ref_s, speed)
        except gr.exceptions.Error as e:
            if use_gpu:
                gr.Warning(str(e))
                gr.Info('Cambiando a CPU')
                audio = models[False](ps, ref_s, speed)
            else:
                raise gr.Error(e)
        yield 24000, audio.numpy()
        if first:
            first = False
            yield 24000, torch.zeros(1).numpy()

with open('es.txt', 'r') as r:
    random_quotes = [line.strip() for line in r]

def get_random_quote():
    return random.choice(random_quotes)

def get_gatsby():
    with open('gatsby5k.md', 'r') as r:
        return r.read().strip()

def get_frankenstein():
    with open('frankenstein5k.md', 'r') as r:
        return r.read().strip()

CHOICES = {
'🇪🇸 🚺 Dora ❤️': 'ef_dora',
'🇪🇸 🚹 Alex': 'em_alex',
'🇪🇸 🚹 Santa': 'em_santa',
}
for v in CHOICES.values():
    pipelines[v[0]].load_voice(v)

TOKEN_NOTE = '''
💡 Ajusta la pronunciación con la sintaxis de enlace de Markdown y /barras diagonales/ así `[Kokoro](/kˈOkəɹO/)`

💬 Para ajustar la entonación, usa puntuación `;:,.!?—…"()“”` o estrés `ˈ` y `ˌ`

⬇️ Disminuye el estrés `[1 nivel](-1)` o `[2 niveles](-2)`

⬆️ Incrementa un nivel `[o](+2)` 2 niveles (solo funciona en palabras menos estresadas, usualmente cortas)
'''

with gr.Blocks() as generate_tab:
    out_audio = gr.Audio(label='Audio resultante', interactive=False, streaming=False, autoplay=True)
    generate_btn = gr.Button('Generar', variant='primary')
    with gr.Accordion('Tokens generados', open=True):
        out_ps = gr.Textbox(interactive=False, show_label=False, info='Tokens usados para generar el audio, contexto de máximo 510.')
        tokenize_btn = gr.Button('Tokenizar', variant='secondary')
        gr.Markdown(TOKEN_NOTE)
        predict_btn = gr.Button('Predecir', variant='secondary', visible=False)

STREAM_NOTE = ['⚠️ Gradio tiene un bug que puede no generar ningún audio la primera vez que hagas clic en `Stream`.']
if CHAR_LIMIT is not None:
    STREAM_NOTE.append(f'✂️ Cada stream se limita a {CHAR_LIMIT} caracteres.')
    STREAM_NOTE.append('🚀 ¿Quieres más caracteres? Puedes [usar Kokoro directamente](https://huggingface.co/hexgrad/Kokoro-82M#usage) o duplicar este espacio:')
STREAM_NOTE = '\n\n'.join(STREAM_NOTE)

with gr.Blocks() as stream_tab:
    out_stream = gr.Audio(label='Stream de audio generado', interactive=False, streaming=True, autoplay=True)
    with gr.Row():
        stream_btn = gr.Button('Stream', variant='primary')
        stop_btn = gr.Button('Detener', variant='stop')
    with gr.Accordion('Nota', open=True):
        gr.Markdown(STREAM_NOTE)
        gr.DuplicateButton()

BANNER_TEXT = '''
[***Kokoro*** **es un modelo de TTS de peso abierto con 82 millones de parámetros.**](https://huggingface.co/hexgrad/Kokoro-82M)

Este demo solo muestra español, puedes encontrar el [original](https://huggingface.co/spaces/hexgrad/Kokoro-TTS) o usarlo directamente para contar con otros idiomas.
'''
API_OPEN = os.getenv('SPACE_ID') != 'hexgrad/Kokoro-TTS'
API_NAME = None if API_OPEN else False
with gr.Blocks() as app:
    with gr.Row():
        gr.Markdown(BANNER_TEXT, container=True)
    with gr.Row():
        with gr.Column():
            text = gr.Textbox(label='Texto a leer', info=f"Máximo ~500 caracteres para «generar», o {'∞' if CHAR_LIMIT is None else CHAR_LIMIT} caracteres usando «Stream»")
            with gr.Row():
                voice = gr.Dropdown(list(CHOICES.items()), value='ef_dora', label='Voz', info='La calidad y disponibilidad varían por idioma')
                use_gpu = gr.Dropdown(
                    [('ZeroGPU 🚀', True), ('CPU 🐌', False)],
                    value=CUDA_AVAILABLE,
                    label='Hardware',
                    info='La GPU usualmente es más rápida, pero tiene quota de uso',
                    interactive=CUDA_AVAILABLE
                )
            speed = gr.Slider(minimum=0.5, maximum=2, value=1, step=0.1, label='Velocidad')
            random_btn = gr.Button('🎲 Cita aleatoria 💬', variant='secondary')
            with gr.Row():
                gatsby_btn = gr.Button('🥂 Gatsby 📕', variant='secondary')
                frankenstein_btn = gr.Button('💀 Frankenstein 📗', variant='secondary')
        with gr.Column():
            gr.TabbedInterface([generate_tab, stream_tab], ['Generar', 'Stream'])
    random_btn.click(fn=get_random_quote, inputs=[], outputs=[text], api_name=API_NAME)
    gatsby_btn.click(fn=get_gatsby, inputs=[], outputs=[text], api_name=API_NAME)
    frankenstein_btn.click(fn=get_frankenstein, inputs=[], outputs=[text], api_name=API_NAME)
    generate_btn.click(fn=generate_first, inputs=[text, voice, speed, use_gpu], outputs=[out_audio, out_ps], api_name=API_NAME)
    tokenize_btn.click(fn=tokenize_first, inputs=[text, voice], outputs=[out_ps], api_name=API_NAME)
    stream_event = stream_btn.click(fn=generate_all, inputs=[text, voice, speed, use_gpu], outputs=[out_stream], api_name=API_NAME)
    stop_btn.click(fn=None, cancels=stream_event)
    predict_btn.click(fn=predict, inputs=[text, voice, speed], outputs=[out_audio], api_name=API_NAME)

if __name__ == '__main__':
    app.queue(api_open=API_OPEN).launch(show_api=API_OPEN, ssr_mode=True)