Spaces:
Runtime error
Runtime error
main.py
Browse files
main.py
ADDED
@@ -0,0 +1,137 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import io
|
2 |
+
import json
|
3 |
+
import os
|
4 |
+
from pathlib import Path
|
5 |
+
|
6 |
+
import librosa
|
7 |
+
import numpy as np
|
8 |
+
import torch
|
9 |
+
import soundfile as sf
|
10 |
+
from demucs.apply import apply_model
|
11 |
+
from demucs.pretrained import DEFAULT_MODEL, get_model
|
12 |
+
from fastapi import FastAPI, UploadFile, File
|
13 |
+
from huggingface_hub import hf_hub_download, list_repo_files
|
14 |
+
from starlette.responses import StreamingResponse
|
15 |
+
|
16 |
+
from so_vits_svc_fork.hparams import HParams
|
17 |
+
from so_vits_svc_fork.inference.core import Svc
|
18 |
+
|
19 |
+
app = FastAPI()
|
20 |
+
|
21 |
+
###################################################################
|
22 |
+
# REPLACE THESE VALUES TO CHANGE THE MODEL REPO/CKPT NAME/SETTINGS
|
23 |
+
###################################################################
|
24 |
+
# The Hugging Face Hub repo ID
|
25 |
+
repo_id = "vettorazi/vettorazi"
|
26 |
+
|
27 |
+
# If None, Uses latest ckpt in the repo
|
28 |
+
ckpt_name = None
|
29 |
+
|
30 |
+
# If None, Uses "kmeans.pt" if it exists in the repo
|
31 |
+
cluster_model_name = None
|
32 |
+
|
33 |
+
# Set the default f0 type to use - use the one it was trained on.
|
34 |
+
# The default for so-vits-svc-fork is "dio".
|
35 |
+
# Options: "crepe", "crepe-tiny", "parselmouth", "dio", "harvest"
|
36 |
+
default_f0_method = "crepe"
|
37 |
+
|
38 |
+
# The default ratio of cluster inference to SVC inference.
|
39 |
+
# If cluster_model_name is not found in the repo, this is set to 0.
|
40 |
+
default_cluster_infer_ratio = 0.5
|
41 |
+
|
42 |
+
# Limit on duration of audio at inference time. increase if you can
|
43 |
+
# In this parent app, we set the limit with an env var to 30 seconds
|
44 |
+
# If you didnt set env var + you go OOM try changing 9e9 to <=300ish
|
45 |
+
duration_limit = int(os.environ.get("MAX_DURATION_SECONDS", 9e9))
|
46 |
+
###################################################################
|
47 |
+
|
48 |
+
if ckpt_name is None:
|
49 |
+
latest_id = sorted(
|
50 |
+
[
|
51 |
+
int(Path(x).stem.split("_")[1])
|
52 |
+
for x in list_repo_files(repo_id)
|
53 |
+
if x.startswith("G_") and x.endswith(".pth")
|
54 |
+
]
|
55 |
+
)[-1]
|
56 |
+
ckpt_name = f"G_{latest_id}.pth"
|
57 |
+
|
58 |
+
cluster_model_name = cluster_model_name or "kmeans.pt"
|
59 |
+
if cluster_model_name in list_repo_files(repo_id):
|
60 |
+
cluster_model_path = hf_hub_download(repo_id, cluster_model_name)
|
61 |
+
else:
|
62 |
+
cluster_model_path = None
|
63 |
+
default_cluster_infer_ratio = default_cluster_infer_ratio if cluster_model_path else 0
|
64 |
+
|
65 |
+
generator_path = hf_hub_download(repo_id, ckpt_name)
|
66 |
+
config_path = hf_hub_download(repo_id, "config.json")
|
67 |
+
hparams = HParams(**json.loads(Path(config_path).read_text()))
|
68 |
+
speakers = list(hparams.spk.keys())
|
69 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
70 |
+
model = Svc(net_g_path=generator_path, config_path=config_path, device=device, cluster_model_path=cluster_model_path)
|
71 |
+
demucs_model = get_model(DEFAULT_MODEL)
|
72 |
+
|
73 |
+
|
74 |
+
def predict(
|
75 |
+
speaker,
|
76 |
+
audio,
|
77 |
+
transpose: int = 0,
|
78 |
+
auto_predict_f0: bool = False,
|
79 |
+
cluster_infer_ratio: float = 0,
|
80 |
+
noise_scale: float = 0.4,
|
81 |
+
f0_method: str = "crepe",
|
82 |
+
db_thresh: int = -40,
|
83 |
+
pad_seconds: float = 0.5,
|
84 |
+
chunk_seconds: float = 0.5,
|
85 |
+
absolute_thresh: bool = False,
|
86 |
+
):
|
87 |
+
audio, _ = librosa.load(audio, sr=model.target_sample, duration=duration_limit)
|
88 |
+
audio = librosa.util.normalize(audio)
|
89 |
+
out = model.predict(
|
90 |
+
audio,
|
91 |
+
speaker,
|
92 |
+
transpose=transpose,
|
93 |
+
auto_predict_f0=auto_predict_f0,
|
94 |
+
cluster_infer_ratio=cluster_infer_ratio,
|
95 |
+
noise_scale=noise_scale,
|
96 |
+
f0_method=f0_method,
|
97 |
+
db_thresh=db_thresh,
|
98 |
+
pad_seconds=pad_seconds,
|
99 |
+
chunk_seconds=chunk_seconds,
|
100 |
+
absolute_thresh=absolute_thresh,
|
101 |
+
)
|
102 |
+
return model.target_sample, out
|
103 |
+
|
104 |
+
|
105 |
+
@app.post("/voice_cloning/")
|
106 |
+
async def voice_cloning(
|
107 |
+
speaker: str,
|
108 |
+
file: UploadFile = File(...),
|
109 |
+
transpose: int = 0,
|
110 |
+
auto_predict_f0: bool = False,
|
111 |
+
cluster_infer_ratio: float = default_cluster_infer_ratio,
|
112 |
+
noise_scale: float = 0.4,
|
113 |
+
f0_method: str = default_f0_method,
|
114 |
+
):
|
115 |
+
# Process the audio file
|
116 |
+
audio_bytes = await file.read()
|
117 |
+
with io.BytesIO(audio_bytes) as audio_io:
|
118 |
+
sample_rate, audio_data = predict(
|
119 |
+
speaker,
|
120 |
+
audio_io,
|
121 |
+
transpose=transpose,
|
122 |
+
auto_predict_f0=auto_predict_f0,
|
123 |
+
cluster_infer_ratio=cluster_infer_ratio,
|
124 |
+
noise_scale=noise_scale,
|
125 |
+
f0_method=f0_method
|
126 |
+
)
|
127 |
+
# Create a BytesIO object to hold the audio data
|
128 |
+
audio_byte_stream = io.BytesIO()
|
129 |
+
# Write the audio data to the BytesIO object
|
130 |
+
sf.write(audio_byte_stream, audio_data, sample_rate, format="wav")
|
131 |
+
# Create a StreamingResponse to return the audio
|
132 |
+
return StreamingResponse(io.BytesIO(audio_byte_stream.getvalue()), media_type="audio/wav")
|
133 |
+
|
134 |
+
|
135 |
+
if __name__ == "__main__":
|
136 |
+
import uvicorn
|
137 |
+
uvicorn.run(app, host="0.0.0.0", port=8000)
|