app.py

import os
import tempfile
import warnings
warnings.filterwarnings("ignore")

import gradio as gr
import numpy as np
import soundfile as sf
import librosa

from huggingface_hub import snapshot_download

# ------------------------------
# Model bootstrap
# ------------------------------
MODEL_DIR = os.path.join(os.getcwd(), "models")
OPENVOICE_REPO = "myshell-ai/OpenVoiceV2"

os.makedirs(MODEL_DIR, exist_ok=True)

# Lazy import to speed up Space boot
_openvoice_loaded = False
_tone_converter = None
_content_extractor = None

_demucs_model = None

def _ensure_openvoice():
    global _openvoice_loaded, _tone_converter, _content_extractor
    if _openvoice_loaded:
        return
    # Download model snapshots into ./models/openvoice
    local_dir = snapshot_download(repo_id=OPENVOICE_REPO, local_dir=os.path.join(MODEL_DIR, "openvoice"), local_dir_use_symlinks=False)

    # OpenVoice v2 layout ships python modules; import after download
    import sys
    if local_dir not in sys.path:
        sys.path.append(local_dir)

    # Import OpenVoice components
    try:
        from openvoice import se_extractor
        from openvoice.api import ToneColorConverter, ContentVec
    except Exception:
        # Fallback to module paths used in some snapshots
        from tone_color_converter.api import ToneColorConverter
        from contentvec.api import ContentVec
        from se_extractor import se_extractor

    # Init content extractor (HuBERT-like)
    content_ckpt = os.path.join(local_dir, "checkpoints", "contentvec", "checkpoint.pth")
    _content_extractor = ContentVec(content_ckpt)

    # Init tone color converter
    tcc_ckpt = os.path.join(local_dir, "checkpoints", "tone_color_converter", "checkpoint.pth")
    _tone_converter = ToneColorConverter(tcc_ckpt, device=os.environ.get("DEVICE", "cuda" if gr.cuda.is_available() else "cpu"))

    _openvoice_loaded = True


def _ensure_demucs():
    global _demucs_model
    if _demucs_model is not None:
        return
    from demucs.apply import apply_model
    from demucs.pretrained import get_model
    from demucs.audio import AudioFile
    _demucs_model = {
        "apply_model": apply_model,
        "get_model": get_model,
        "AudioFile": AudioFile,
    }


def separate_vocals(wav_path, stem="vocals"):
    """Return path to separated vocals and accompaniment using htdemucs."""
    _ensure_demucs()
    apply_model = _demucs_model["apply_model"]
    get_model = _demucs_model["get_model"]
    AudioFile = _demucs_model["AudioFile"]

    model = get_model(name="htdemucs")
    model.cpu()

    with AudioFile(wav_path).read(streams=0, samplerate=44100, channels=2) as mix:
        ref = mix
        out = apply_model(model, ref, shifts=1, split=True, overlap=0.25)
        sources = {name: out[idx] for idx, name in enumerate(model.sources)}

    # Save stems
    base = os.path.splitext(os.path.basename(wav_path))[0]
    out_dir = tempfile.mkdtemp(prefix="stems_")
    vocal_path = os.path.join(out_dir, f"{base}_vocals.wav")
    inst_path = os.path.join(out_dir, f"{base}_inst.wav")

    sf.write(vocal_path, sources["vocals"].T, 44100)
    # Combine other stems for instrumental
    inst = sum([v for k, v in sources.items() if k != "vocals"]) / (len(model.sources) - 1)
    sf.write(inst_path, inst.T, 44100)
    return vocal_path, inst_path


def load_audio(x, sr=44100, mono=True):
    y, _sr = librosa.load(x, sr=sr, mono=mono)
    return y, sr


def save_audio(y, sr):
    path = tempfile.mktemp(suffix=".wav")
    sf.write(path, y, sr)
    return path


def match_length(a, b):
    # Pad/trim a to match length of b
    if len(a) < len(b):
        a = np.pad(a, (0, len(b)-len(a)))
    else:
        a = a[:len(b)]
    return a


def convert_voice(reference_wav, source_vocal_wav, style_strength=0.8, pitch_shift=0.0, formant_shift=0.0):
    _ensure_openvoice()

    # Load audio
    ref, sr = load_audio(reference_wav, sr=16000, mono=True)
    src, _ = load_audio(source_vocal_wav, sr=16000, mono=True)

    # Extract content features from source
    content = _content_extractor.extract(src, sr)

    # Extract speaker embedding / tone color from reference
    # OpenVoice ships an SE (speaker encoder) util; we mimic via API if exposed.
    try:
        from openvoice import se_extractor
        se = se_extractor.get_se(reference_wav, device=_tone_converter.device)
    except Exception:
        # Some snapshots provide a function name get_se_wav
        from se_extractor import get_se
        se = get_se(reference_wav)

    # Run tone color conversion
    converted = _tone_converter.convert(content, se, style_strength=style_strength)

    y = converted

    # Optional pitch & formant adjustments (light touch)
    if abs(pitch_shift) > 1e-3:
        y = librosa.effects.pitch_shift(y.astype(np.float32), 16000, n_steps=pitch_shift)
    if abs(formant_shift) > 1e-3:
        # crude formant-esque EQ tilt using shelving filter via librosa
        import scipy.signal as sps
        w = 2 * np.pi * 1500 / 16000
        b, a = sps.iirfilter(2, Wn=w/np.pi, btype='high', ftype='but