app.py

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

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

try:
    import phonemizer
    global_phonemizer_en = phonemizer.backend.EspeakBackend(language='en-us', preserve_punctuation=True, with_stress=True)
    PHONEMIZER_AVAILABLE = True
    global_phonemizer_en = None
except ImportError:
    PHONEMIZER_AVAILABLE = False

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 'ab'}
pipelines['a'].g2p.lexicon.golds['kokoro'] = 'kˈOkəɹO'
pipelines['b'].g2p.lexicon.golds['kokoro'] = 'kˈQkəɹQ'

# Language mappings similar to StyleTTS
LANG_NAMES = {
    'en-us': 'english',
    'cs': 'czech',
    'da': 'danish',
    'nl': 'dutch',
    'et': 'estonian',
    'fi': 'finnish',
    'fr': 'french',
    'de': 'german',
    'el': 'greek',
    'it': 'italian',
    'no': 'norwegian',
    'pl': 'polish',
    'pt': 'portuguese',
    'ru': 'russian',
    'sl': 'slovene',
    'es': 'spanish',
    'sv': 'swedish',
    'tr': 'turkish',
}

def text_to_ipa(text, lang='en-us'):
    """Convert text to IPA using phonemizer or return original text"""
    if not PHONEMIZER_AVAILABLE:
        return text

    try:
        # Handle IPA sections within brackets
        regex = r"\([^\]]*\)[[^\]]*\]"
        ipa_sections = re.findall(regex, text)
        print(text)
        text = re.sub(regex, '()[]', text)
        print(text)

        if lang == 'jb':
            # Lojban language
            import lojban
            ps = f'[{text}](/'+ lojban.lojban2ipa(text, 'vits') +'/)'
        elif lang in LANG_NAMES:
            local_phonemizer = phonemizer.backend.EspeakBackend(language=lang, preserve_punctuation=True, with_stress=True)
            ps = local_phonemizer.phonemize([text])
            ps = f'[{text}](/'+ ps[0] +'/)'
        else:
            ps = text
        print(ps)

        # Add back IPA sections
        for ipa in ipa_sections:
            ps = ps.replace('( )[ ]', ipa, 1)
        print(ps)

        return ps
    except Exception as e:
        print(f"Phonemizer error: {e}")
        return text

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

def generate_first(text, voice='af_heart', speed=1, use_gpu=CUDA_AVAILABLE, lang='en-us'):
    text = text if CHAR_LIMIT is None else text.strip()[:CHAR_LIMIT]

    # Convert text to IPA if not English
    if lang != 'en-us':
        text = text_to_ipa(text, lang)

    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('Retrying with CPU. To avoid this error, change Hardware to 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='af_heart', speed=1):
    """ Convert the text into speech using Kokoro American English and British English voice models.

        Args:
            text: string; accepts IPA within ()[] brackets
            voice: Literal['af_heart', 'af_bella', 'af_nicole', 'af_aoede', 'af_kore', 'af_sarah', 'af_nova', 'af_sky', 'af_alloy', 'af_jessica', 'af_river', 'am_michael', 'am_fenrir', 'am_puck', 'am_echo', 'am_eric', 'am_liam', 'am_onyx', 'am_santa', 'am_adam', 'bf_emma', 'bf_isabella', 'bf_alice', 'bf_lily', 'bm_george', 'bm_fable', 'bm_lewis', 'bm_daniel']; voice model
            lang: Literal['en-us', 'cs', 'da', 'nl', 'et', 'fi', 'fr', 'de', 'el', 'it', 'no', 'pl', 'pt', 'ru', 'sl', 'es', 'sv', 'tr', 'jb']; ISO 639-1 code for the text language; 'jb' is a valid code for Lojban
            speed: talkback speed; 0.5-2

        Returns: Tuple of (output_audio_path, ipa_results) where output_audio_path is the filepath of output audio
    """
    return generate_first(text, voice, speed, use_gpu=False)[0]

def tokenize_first(text, voice='af_heart', lang='en-us'):
    text = text if CHAR_LIMIT is None else text.strip()[:CHAR_LIMIT]

    # Convert text to IPA if not English or if language is specified
    if lang != 'en-us':
        text = text_to_ipa(text, lang)

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

def generate_all(text, voice='af_heart', speed=1, use_gpu=CUDA_AVAILABLE, lang='en-us'):
    text = text if CHAR_LIMIT is None else text.strip()[:CHAR_LIMIT]

    # Convert text to IPA if not English or if language is specified
    if lang != 'en-us':
        text = text_to_ipa(text, lang)

    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('Switching to 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('en.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 = {
'🇺🇸 🚺 Heart ❤️': 'af_heart',
'🇺🇸 🚺 Bella 🔥': 'af_bella',
'🇺🇸 🚺 Nicole 🎧': 'af_nicole',
'🇺🇸 🚺 Aoede': 'af_aoede',
'🇺🇸 🚺 Kore': 'af_kore',
'🇺🇸 🚺 Sarah': 'af_sarah',
'🇺🇸 🚺 Nova': 'af_nova',
'🇺🇸 🚺 Sky': 'af_sky',
'🇺🇸 🚺 Alloy': 'af_alloy',
'🇺🇸 🚺 Jessica': 'af_jessica',
'🇺🇸 🚺 River': 'af_river',
'🇺🇸 🚹 Michael': 'am_michael',
'🇺🇸 🚹 Fenrir': 'am_fenrir',
'🇺🇸 🚹 Puck': 'am_puck',
'🇺🇸 🚹 Echo': 'am_echo',
'🇺🇸 🚹 Eric': 'am_eric',
'🇺🇸 🚹 Liam': 'am_liam',
'🇺🇸 🚹 Onyx': 'am_onyx',
'🇺🇸 🚹 Santa': 'am_santa',
'🇺🇸 🚹 Adam': 'am_adam',
'🇬🇧 🚺 Emma': 'bf_emma',
'🇬🇧 🚺 Isabella': 'bf_isabella',
'🇬🇧 🚺 Alice': 'bf_alice',
'🇬🇧 🚺 Lily': 'bf_lily',
'🇬🇧 🚹 George': 'bm_george',
'🇬🇧 🚹 Fable': 'bm_fable',
'🇬🇧 🚹 Lewis': 'bm_lewis',
'🇬🇧 🚹 Daniel': 'bm_daniel',
}
for v in CHOICES.values():
    pipelines[v[0]].load_voice(v)

TOKEN_NOTE = '''
💡 Customize pronunciation with Markdown link syntax and /slashes/ like `[Kokoro](/kˈOkəɹO/)`

💬 To adjust intonation, try punctuation `;:,.!?—…"()“”` or stress `ˈ` and `ˌ`

⬇️ Lower stress `[1 level](-1)` or `[2 levels](-2)`

⬆️ Raise stress 1 level `[or](+2)` 2 levels (only works on less stressed, usually short words)
'''

with gr.Blocks() as generate_tab:
    out_audio = gr.Audio(label='Output Audio', interactive=False, streaming=False, autoplay=True)
    generate_btn = gr.Button('Generate', variant='primary')
    with gr.Accordion('Output Tokens', open=True):
        out_ps = gr.Textbox(interactive=False, show_label=False, info='Tokens used to generate the audio, up to 510 context length.')
        tokenize_btn = gr.Button('Tokenize', variant='secondary')
        gr.Markdown(TOKEN_NOTE)
        predict_btn = gr.Button('Predict', variant='secondary', visible=False)

STREAM_NOTE = ['⚠️ There is an unknown Gradio bug that might yield no audio the first time you click `Stream`.']
if CHAR_LIMIT is not None:
    STREAM_NOTE.append(f'✂️ Each stream is capped at {CHAR_LIMIT} characters.')
    STREAM_NOTE.append('🚀 Want more characters? You can [use Kokoro directly](https://huggingface.co/hexgrad/Kokoro-82M#usage) or duplicate this space:')
STREAM_NOTE = '\n\n'.join(STREAM_NOTE)

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

BANNER_TEXT = '''
[***Kokoro*** **is an open-weight TTS model with 82 million parameters.**](https://huggingface.co/hexgrad/Kokoro-82M)

This demo uses native US English and British English speakers. But also supports multiple languages using G2P and phonemizer. Select your language below!
'''

API_OPEN = os.getenv('SPACE_ID') != 'hexgrad/Kokoro-TTS'
API_NAME = None if API_OPEN else False

# Language choices for the dropdown
LANGUAGE_CHOICES = [
    ['English (US)', 'en-us'],
    ['Lojban', 'jb'],
    ['Czech (Non-native)', 'cs'],
    ['Danish (Non-native)', 'da'],
    ['Dutch (Non-native)', 'nl'],
    ['Estonian (Non-native)', 'et'],
    ['Finnish (Non-native)', 'fi'],
    ['French (Non-native)', 'fr'],
    ['German (Non-native)', 'de'],
    ['Greek (Non-native)', 'el'],
    ['Italian (Non-native)', 'it'],
    ['Norwegian (Non-native)', 'no'],
    ['Polish (Non-native)', 'pl'],
    ['Portuguese (Non-native)', 'pt'],
    ['Russian (Non-native)', 'ru'],
    ['Slovene (Non-native)', 'sl'],
    ['Spanish (Non-native)', 'es'],
    ['Swedish (Non-native)', 'sv'],
    ['Turkish (Non-native)', 'tr'],
]

with gr.Blocks() as app:
    with gr.Row():
        gr.Markdown(BANNER_TEXT, container=True)
    with gr.Row():
        with gr.Column():
            text = gr.Textbox(label='Input Text', info=f"Up to ~500 characters per Generate, or {'∞' if CHAR_LIMIT is None else CHAR_LIMIT} characters per Stream. Supports IPA within (<text>)[<IPA>] parethesis and brackets.")
            with gr.Row():
                voice = gr.Dropdown(list(CHOICES.items()), value='af_heart', label='Voice', info='Quality and availability vary by language')
                use_gpu = gr.Dropdown(
                    [('ZeroGPU 🚀', True), ('CPU 🐌', False)],
                    value=CUDA_AVAILABLE,
                    label='Hardware',
                    info='GPU is usually faster, but has a usage quota',
                    interactive=CUDA_AVAILABLE
                )
            with gr.Row():
                lang = gr.Dropdown(
                    LANGUAGE_CHOICES,
                    value='en-us',
                    label="Language",
                    info="Select language for G2P processing"
                )
                speed = gr.Slider(minimum=0.5, maximum=2, value=1, step=0.1, label='Speed')
            random_btn = gr.Button('🎲 Random Quote 💬', 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], ['Generate', '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, lang], outputs=[out_audio, out_ps], api_name=API_NAME)
    tokenize_btn.click(fn=tokenize_first, inputs=[text, voice, lang], outputs=[out_ps], api_name=API_NAME)
    stream_event = stream_btn.click(fn=generate_all, inputs=[text, voice, speed, use_gpu, lang], 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, mcp_server=API_OPEN)