app.py

import io
import gradio as gr
import librosa
import numpy as np
import soundfile
from inference.infer_tool import Svc

import os
def list_files_tree(directory, indent=""):
    items = os.listdir(directory)
    for i, item in enumerate(items):
        prefix = "└── " if i == len(items) - 1 else "├── "
        print(indent + prefix + item)
        item_path = os.path.join(directory, item)
        if os.path.isdir(item_path):
            next_indent = indent + ("    " if i == len(items) - 1 else "│   ")
            list_files_tree(item_path, next_indent)

from huggingface_hub import snapshot_download
print("Models...")
models_id = """None1145/So-VITS-SVC-Vulpisfoglia
None1145/So-VITS-SVC-Lappland
None1145/So-VITS-SVC-Lappland-the-Decadenza
None1145/So-VITS-SVC-Rosmontis"""
for model_id in models_id.split("\n"):
    if model_id in ["", " "]:
        break
    print(f"{model_id}...")
    snapshot_download(repo_id=model_id, local_dir=f"./Models/{model_id}")
    print(f"{model_id}!!!")
print("Models!!!")
print("PretrainedModels...")
base_model_id = "None1145/So-VITS-SVC-Base"
snapshot_download(repo_id=base_model_id, local_dir=f"./PretrainedModels/{base_model_id}")
print("PretrainedModels!!!")
list_files_tree("./")

import re
models_info = {}
models_folder_path = "./Models/None1145"
folder_names = [name for name in os.listdir(models_folder_path) if os.path.isdir(os.path.join(models_folder_path, name))]
for folder_name in folder_names:
    speaker = folder_name[12:]
    pattern = re.compile(r"G_(\d+)\.pth$")
    max_value = -1
    max_file = None
    models_path = f"{models_folder_path}/{folder_name}/Models"
    config_path = f"{models_folder_path}/{folder_name}/Configs"
    for filename in os.listdir(models_path):
        match = pattern.search(filename)
        if match:
            value = int(match.group(1))
            if value > max_value:
                max_value = value
                max_file = filename
    models_info[speaker] = {}
    models_info[speaker]["model"] = f"{models_path}/{max_file}"
    models_info[speaker]["config"] = f"{config_path}/config.json"
    if os.path.exists(f"{models_path}/feature_and_index.pkl"):
        models_info[speaker]["cluster"] = f"{models_path}/feature_and_index.pkl"
        models_info[speaker]["feature_retrieval"] = True
    elif os.path.exists(f"{models_path}/kmeans_10000.pt"):
        models_info[speaker]["cluster"] = f"{models_path}/kmeans_10000.pt"
        models_info[speaker]["feature_retrieval"] = False
    else:
        models_info[speaker]["cluster"] = "logs/44k/kmeans_10000.pt"
        models_info[speaker]["feature_retrieval"] = False
speakers = list(models_info.keys())
print(models_info)
print(speakers)

def load(speaker):
    return Svc(models_info[speaker]["model"], models_info[speaker]["config"], cluster_model_path=models_info[speaker]["cluster"], feature_retrieval=models_info[speaker]["feature_retrieval"])

def vc_fn(speaker, input_audio, vc_transform, auto_f0,cluster_ratio, slice_db, noise_scale):
    model = load(speaker)
    if input_audio is None:
        return "You need to upload an audio", None
    sampling_rate, audio = input_audio
    duration = audio.shape[0] / sampling_rate
    audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
    if len(audio.shape) > 1:
        audio = librosa.to_mono(audio.transpose(1, 0))
    if sampling_rate != 16000:
        audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
    print(audio.shape)
    out_wav_path = "temp.wav"
    soundfile.write(out_wav_path, audio, 16000, format="wav")
    print( cluster_ratio, auto_f0, noise_scale)
    _audio = model.slice_inference(out_wav_path, speaker, vc_transform, slice_db, cluster_ratio, auto_f0, noise_scale)
    return "Success", (44100, _audio)

app = gr.Blocks()
with app:
    with gr.Tabs():
        for speaker in speakers:
            with gr.TabItem(speaker):
                with gr.Row():
                    gr.Markdown(
                        '<div align="center">'
                        f'<a><strong>{speaker}</strong></a>'
                        '</div>')
                speaker = gr.Textbox(label="Speaker", value=speaker)
                vc_input3 = gr.Audio(label="Upload Audio")
                vc_transform = gr.Number(label="Pitch Shift (integer, can be positive or negative, number of semitones, raising an octave is +12)", value=0)
                cluster_ratio = gr.Number(label="Cluster Model Mixing Ratio (0-1): Defaults to 0 (clustering disabled). Improves timbre similarity but may reduce articulation clarity. Recommended value: ~0.5 if used", value=0)
                auto_f0 = gr.Checkbox(label="Auto f0 Prediction: Works better with the cluster model for f0 prediction but disables the pitch shift feature. (For voice conversion only; do not enable this for singing voices, as it will result in extreme off-pitch issues)", value=False)
                slice_db = gr.Number(label="Slicing Threshold", value=-40)
                noise_scale = gr.Number(label="noise_scale", value=0.4)
                vc_submit = gr.Button("Convert", variant="primary")
                vc_output1 = gr.Textbox(label="Output Message")
                vc_output2 = gr.Audio(label="Output Audio")
                vc_submit.click(vc_fn, [speaker, vc_input3, vc_transform,auto_f0,cluster_ratio, slice_db, noise_scale], [vc_output1, vc_output2])
    app.launch()