app.py

from pathlib import Path
import torch
from so_vits_svc_fork.hparams import HParams
import json
import gradio as gr
import librosa
import numpy as np

##########################################################
# REPLACE THESE VALUES TO CHANGE THE MODEL REPO/CKPT NAME
##########################################################
repo_id = "dog/theovon"
ckpt_name = None # or specify a ckpt. ex. "G_1257.pth"
##########################################################

# Figure out the latest generator by taking highest value one.
# Ex. if the repo has: G_0.pth, G_100.pth, G_200.pth, we'd use G_200.pth
if ckpt_name is None:
    latest_id = sorted(
        [
            int(Path(x).stem.split("_")[1])
            for x in list_repo_files(repo_id)
            if x.startswith("G_") and x.endswith(".pth")
        ]
    )[-1]
    ckpt_name = f"G_{latest_id}.pth"

generator_path = hf_hub_download(repo_id, ckpt_name)
config_path = hf_hub_download(repo_id, "config.json")
hparams = HParams(**json.loads(Path(config_path).read_text()))
speakers = list(hparams.spk.keys())
device = "cuda" if torch.cuda.is_available() else "cpu"
model = Svc(net_g_path=generator_path, config_path=config_path, device=device, cluster_model_path=None)


def predict(
    speaker,
    audio,
    transpose: int = 0,
    auto_predict_f0: bool = False,
    cluster_infer_ratio: float = 0,
    noise_scale: float = 0.4,
    f0_method: str = "crepe",
    db_thresh: int = -40,
    pad_seconds: float = 0.5,
    chunk_seconds: float = 0.5,
    absolute_thresh: bool = False,
):
    audio, _ = librosa.load(audio, sr=model.target_sample)
    audio = model.infer_silence(
        audio.astype(np.float32),
        speaker=speaker,
        transpose=transpose,
        auto_predict_f0=auto_predict_f0,
        cluster_infer_ratio=cluster_infer_ratio,
        noise_scale=noise_scale,
        f0_method=f0_method,
        db_thresh=db_thresh,
        pad_seconds=pad_seconds,
        chunk_seconds=chunk_seconds,
        absolute_thresh=absolute_thresh,
    )
    return model.target_sample, audio


interface = gr.Interface(
    predict,
    inputs=[
        gr.Dropdown(speakers, value=speakers[0], label="Target Speaker"),
        gr.Audio(type="filepath", source="microphone", label="Source Audio"),
        gr.Slider(-12, 12, value=0, step=1, label="Transpose (Semitones)"),
        gr.Checkbox(False, label="Auto Predict F0"),
        gr.Slider(0.0, 1.0, value=0.0, step=0.1, label='cluster infer ratio'),
        gr.Slider(0.0, 1.0, value=0.4, step=0.1, label="noise scale"),
        gr.Dropdown(choices=["crepe", "crepe-tiny", "parselmouth", "dio", "harvest"], value='crepe', label="f0 method"),
    ],
    outputs="audio",
    title="Voice Cloning",
    description=f"""
This app uses models trained with so-vits-svc-fork to clone your voice. Model currently being used is https://hf.co/{repo_id}. To change the model being served, duplicate the space and update the repo_id in `app.py`.
""".strip(),
    article="""
<p style='text-align: center'>
    <a href='https://github.com/voicepaw/so-vits-svc-fork' target='_blank'>Github Repo</a>
</p>
"""
)

if __name__ == '__main__':
    interface.launch()