app.py

"""Maya 2 Native — multilingual streaming TTS for 12 Indic languages + English,
with zero-shot voice cloning. Apache-2.0, 2.5B params, 24 kHz output.

ZeroGPU note: the maya3 engine captures CUDA graphs and allocates pinned host
buffers at init, which requires a real CUDA context — impossible at module
scope under the ZeroGPU hijack. So the engine is built lazily inside the
@spaces.GPU worker (cached in a module global, which persists for the warm
worker process), while the ~5 GB of weights are pre-downloaded to the HF cache
at module scope so worker-side loads are disk-bound.
"""
import os
import time

import spaces  # MUST precede any torch / CUDA import

os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")

# --- module scope: pre-warm the HF cache (pure disk/network, no CUDA) -------
from maya3 import hub

RELEASE = "maya-research/maya2-native"
_paths = hub.resolve(RELEASE)  # downloads model shards + kyutai/mimi + speechbrain ECAPA
CKPT_DIR = _paths["ckpt"]      # release dir; from_pretrained re-resolves offline-first

import gradio as gr
import numpy as np

LANGS = ["hi", "en", "as", "bn", "gu", "kn", "ml", "mr", "or", "pa", "ta", "te"]
LANG_NAMES = {
    "hi": "Hindi", "en": "English", "as": "Assamese", "bn": "Bengali",
    "gu": "Gujarati", "kn": "Kannada", "ml": "Malayalam", "mr": "Marathi",
    "or": "Odia", "pa": "Punjabi", "ta": "Tamil", "te": "Telugu",
}
# 20 named voices shipped with the release (voices/voices.json).
VOICES = [
    "Ahana", "Vaidehi", "Malini", "Charvi", "Jahnavi",
    "Anirban", "Nakul", "Vasant", "Siddhant", "Satvik",
    "Abhay", "Nivedita", "Ishwari", "Ritvik", "Vrinda",
    "Sreekar", "Kayal", "Shaunak", "Akshat", "Tanushree",
]

MAX_CHARS = 1200  # ~1 minute of speech; keeps duration bounded

_tts = None


def _get_tts():
    """Build the engine once per worker process (weights load + CUDA graphs + warmup)."""
    global _tts
    if _tts is None:
        from maya3.api import Maya3TTS

        t0 = time.perf_counter()
        _tts = Maya3TTS.from_pretrained(
            CKPT_DIR,          # release dir; re-resolves (offline-first) + registers named voices
            device="cuda",
            max_seats=1,       # smallest graph grid; demo serves one request at a time
            capacity=1024,     # halves KV cache vs default; still ~80 s of audio per piece
        )
        print(f"[maya2] engine ready in {time.perf_counter() - t0:.1f}s", flush=True)
    return _tts


@spaces.GPU(duration=90)  # cold worker: ~48s engine init + a few s generation; warm: ~2s
def generate(
    text: str,
    lang: str,
    voice: str,
    normalize: bool,
    seed: int,
    ref_audio=None,
    ref_text: str = "",
    ref_lang: str = "en",
    progress=gr.Progress(),
):
    """Synthesize speech from text in one of 12 Indic languages or English.

    Optionally clone a voice from a 3-10 s reference clip: upload the clip,
    give its exact transcript, and the same speaker will read your text —
    even in a different language.

    Returns 24 kHz mono audio.
    """
    text = (text or "").strip()
    if not text:
        raise gr.Error("Please enter some text to synthesize.")
    if len(text) > MAX_CHARS:
        raise gr.Error(f"Text too long ({len(text)} chars); please keep it under {MAX_CHARS}.")

    tts = _get_tts()

    kw = {}
    if ref_audio:
        if not (ref_text or "").strip():
            raise gr.Error("Voice cloning needs the reference clip's transcript (ref_text).")
        kw = dict(ref_audio=ref_audio, ref_text=ref_text, ref_lang=ref_lang)
        voice = None  # engine forbids combining a named voice with ref_audio
    elif voice == "(none)":
        voice = None

    t0 = time.perf_counter()
    try:
        wav, info = tts.generate(text, lang=lang, voice=voice, normalize=normalize, seed=seed,
                                 return_info=True)
    except Exception as e:
        msg = str(e)
        code = getattr(e, "code", None)
        raise gr.Error(f"{code}: {msg}" if code else msg)
    dt = time.perf_counter() - t0
    dur = len(wav) / 24000
    print(f"[maya2] lang={lang} voice={voice or 'clone'} chars={len(text)} "
          f"audio={dur:.1f}s gpu={dt:.1f}s reason={info.get('reason')} frames={info.get('frames')}",
          flush=True)
    if len(wav) == 0:
        raise gr.Error("The model produced no audio — try rephrasing or raising max length.")
    return (24000, wav)


CSS = """
main, .gradio-container {max-width: 900px !important; margin: 0 auto !important;}
.dark .gradio-container {color: var(--body-text-color);}
"""

with gr.Blocks() as demo:
    gr.HTML("""
<div style="display: flex; align-items: center; gap: 12px; justify-content: center; margin-bottom: 4px;">
  <div style="font-size: 44px; line-height: 1;">🗣️</div>
  <div>
    <h1 style="margin: 0; font-size: 28px;">Maya 2 Native</h1>
    <p style="margin: 0; opacity: 0.75;">Multilingual streaming TTS — 12 Indic languages + English, with zero-shot voice cloning</p>
  </div>
</div>
<p style="text-align: center; opacity: 0.7; margin-top: 0;">
2.5B params · Apache-2.0 · 24 kHz audio · <a href="https://huggingface.co/maya-research/maya2-native" target="_blank">model card</a>
</p>
""")

    with gr.Row():
        with gr.Column(scale=3):
            text_in = gr.Textbox(
                label="Text to speak",
                placeholder="Enter text in any of the 12 supported languages…",
                lines=4,
            )
            with gr.Row():
                lang_in = gr.Dropdown(
                    choices=[(f"{LANG_NAMES[c]} ({c})", c) for c in LANGS],
                    value="hi",
                    label="Language",
                    filterable=True,
                )
                voice_in = gr.Dropdown(
                    choices=["(none)"] + VOICES,
                    value="Ritvik",
                    label="Voice",
                    filterable=True,
                    info="Named voices ship with the model; or clone your own below.",
                )
            audio_out = gr.Audio(label="Generated speech", type="filepath")
            btn = gr.Button("Generate speech", variant="primary")
        with gr.Column(scale=2):
            gr.Markdown("### Clone a voice (optional)")
            ref_audio = gr.Audio(
                label="Reference clip (3–10 s)",
                type="filepath",
                sources=["upload", "microphone"],
                format="wav",
            )
            ref_text = gr.Textbox(
                label="Transcript of the reference clip",
                placeholder="What is said in the clip, exactly…",
                lines=2,
            )
            ref_lang = gr.Dropdown(
                choices=[(f"{LANG_NAMES[c]} ({c})", c) for c in LANGS],
                value="hi",
                label="Reference language",
            )
            gr.Markdown(
                "<span style='opacity:0.7'>Upload a clip and its exact transcript; the same speaker "
                "will read your text — even in a different language.</span>"
            )

    with gr.Accordion("Advanced options", open=False):
        normalize = gr.Checkbox(
            value=True,
            label="Normalize text (digits, currency, dates, URLs)",
            info="Recommended: the bundled normaliser expands numbers, dates and symbols into words.",
        )
        seed = gr.Number(
            value=0,
            label="Seed (0 = random)",
            precision=0,
            minimum=0,
            maximum=2**31 - 1,
        )

    gr.Examples(
        examples=[
            ["क्या आपका वर्तमान पता 45, एमजी रोड, अंधेरी ईस्ट, मुंबई 400069 है? यदि हाँ, तो 1 दबाएं।", "hi", "Ritvik"],
            ["The EMI for your home loan is due by 12-10-24; please ensure your account has sufficient funds.", "en", "Siddhant"],
            ["જી બિલકુલ, મને સિસ્ટમમાં દેખાય છે કે તમારું ₹2,500 નું પેમેન્ટ 15મી તારીખે સક્સેસફુલ થઈ ગયું છે.", "gu", "Vaidehi"],
            ["ಹೌದಾ... ಅದು ನಿಜಕ್ಕೂ ಕಷ್ಟ. ನಿಮ್ಮ ಅಂತಿಮ ಬಿಲ್ ₹45,000 ಆಗಿದೆ, ಅದರಲ್ಲಿ ಇನ್ಶೂರೆನ್ಸ್ ಕವರ್ ಆಗಿದೆಯಾ ಅಂತ ಚೆಕ್ ಮಾಡ್ತೀನಿ.", "kn", "Satvik"],
            ["റദ്ദാക്കിയ ടിക്കറ്റിന്റെ റീഫണ്ട് തുകയായ ₹1,800 മൂന്ന് ദിവസത്തിനുള്ളിൽ ലഭിക്കും.", "ml", "Anirban"],
            ["मला 98765 43210 या नंबरवर OTP आला नाही, कृपया पुन्हा resend करा.", "mr", "Charvi"],
            ["ଦଯ଼ାକରି ମନେରଖନ୍ତୁ ଯେ EMI ପ୍ରତ୍ୟେକ ମାସର 5 ତାରିଖରେ କଟିବ ଏବଂ ଏହା ଆପଣଙ୍କର 3ୟ କିସ୍ତି।", "or", "Ahana"],
            ["NEFT ਰਾਹੀਂ ਫੰਡ ਟਰਾਂਸਫਰ 00:30 ਤੋਂ 23:30 ਤੱਕ ਉਪਲਬਧ ਹਨ।", "pa", "Malini"],
            ["கடந்த நிதியாண்டில் வங்கியின் மொத்த லாபம் ₹3.5 கோடி, இதில் ₹45 லட்சம் முதலீட்டாளர்களுக்கு வழங்கப்பட்டது.", "ta", "Abhay"],
            ["సేవింగ్స్ తో పోలిస్తే ఫిక్స్డ్ డిపాజిట్ వడ్డీ నిష్పత్తి 3:2 గా ఉండి, 7.25% ఇస్తుంది.", "te", "Nivedita"],
            ["এন ই এফ টি লেনদেনটো সম্পূৰ্ণ হ'বলৈ প্ৰায় 2 ঘণ্টা 30 মিনিট সময় লাগিব পাৰে।", "as", "Nakul"],
            ["12/08/2023 তারিখে আপনার 1,500 টাকার refund process হয়ে গেছে, status update চেক করুন।", "bn", "Jahnavi"],
        ],
        inputs=[text_in, lang_in, voice_in],
        fn=generate,
        outputs=[audio_out],
        cache_examples=True,
        cache_mode="lazy",
        label="Try the authors' showcase examples (one per supported language)",
    )

    gr.Markdown(
        "<span style='opacity:0.65'>Voice-cloning example: leave the examples alone, upload your own "
        "3–10 s clip in <b>Clone a voice</b>, type its exact transcript, then press Generate. "
        "A bundled reference clip (`samples/hi_ritvik.wav`) is in this Space's files for inspiration.</span>"
    )

    btn.click(
        generate,
        inputs=[text_in, lang_in, voice_in, normalize, seed, ref_audio, ref_text, ref_lang],
        outputs=[audio_out],
        api_name="generate",
        concurrency_limit=1,
    )


if __name__ == "__main__":
    demo.launch(theme=gr.themes.Citrus(), css=CSS, mcp_server=True)