add gradio app
Browse files
app.py
ADDED
@@ -0,0 +1,150 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
## VCTK
|
2 |
+
import torch
|
3 |
+
import os
|
4 |
+
|
5 |
+
import commons
|
6 |
+
import utils
|
7 |
+
from models import SynthesizerTrn
|
8 |
+
from text.symbols import symbols
|
9 |
+
from text import text_to_sequence
|
10 |
+
|
11 |
+
from scipy.io.wavfile import write
|
12 |
+
import gradio as gr
|
13 |
+
|
14 |
+
print("Running GRadio", gr.__version__)
|
15 |
+
|
16 |
+
model_path = "vits2_pytorch/G_390000.pth"
|
17 |
+
config_path = "vits2_pytorch/vits2_vctk_cat_inference.json"
|
18 |
+
|
19 |
+
hps = utils.get_hparams_from_file(config_path)
|
20 |
+
|
21 |
+
if (
|
22 |
+
"use_mel_posterior_encoder" in hps.model.keys()
|
23 |
+
and hps.model.use_mel_posterior_encoder == True
|
24 |
+
):
|
25 |
+
print("Using mel posterior encoder for VITS2")
|
26 |
+
posterior_channels = 80 # vits2
|
27 |
+
hps.data.use_mel_posterior_encoder = True
|
28 |
+
else:
|
29 |
+
print("Using lin posterior encoder for VITS1")
|
30 |
+
posterior_channels = hps.data.filter_length // 2 + 1
|
31 |
+
hps.data.use_mel_posterior_encoder = False
|
32 |
+
|
33 |
+
net_g = SynthesizerTrn(
|
34 |
+
len(symbols),
|
35 |
+
posterior_channels,
|
36 |
+
hps.train.segment_size // hps.data.hop_length,
|
37 |
+
n_speakers=hps.data.n_speakers,
|
38 |
+
**hps.model
|
39 |
+
)
|
40 |
+
_ = net_g.eval()
|
41 |
+
|
42 |
+
_ = utils.load_checkpoint(model_path, net_g, None)
|
43 |
+
|
44 |
+
|
45 |
+
def get_text(text, hps):
|
46 |
+
text_norm = text_to_sequence(text, hps.data.text_cleaners)
|
47 |
+
#text_norm = cleaned_text_to_sequence(text) # if model was trained with text
|
48 |
+
|
49 |
+
if hps.data.add_blank:
|
50 |
+
text_norm = commons.intersperse(text_norm, 0)
|
51 |
+
text_norm = torch.LongTensor(text_norm)
|
52 |
+
return text_norm
|
53 |
+
|
54 |
+
def tts(text:str, speaker_id:int, speed:float, noise_scale:float=0.667, noise_scale_w:float=0.8):
|
55 |
+
|
56 |
+
stn_tst = get_text(text, hps)
|
57 |
+
with torch.no_grad():
|
58 |
+
x_tst = stn_tst.unsqueeze(0)
|
59 |
+
x_tst_lengths = torch.LongTensor([stn_tst.size(0)])
|
60 |
+
sid = torch.LongTensor([speaker_id])
|
61 |
+
waveform = (
|
62 |
+
net_g.infer(
|
63 |
+
x_tst,
|
64 |
+
x_tst_lengths,
|
65 |
+
sid=sid,
|
66 |
+
noise_scale=noise_scale,
|
67 |
+
noise_scale_w=noise_scale_w,
|
68 |
+
length_scale=1/speed,
|
69 |
+
)[0][0, 0]
|
70 |
+
.data.cpu()
|
71 |
+
.float()
|
72 |
+
.numpy()
|
73 |
+
)
|
74 |
+
|
75 |
+
return gr.make_waveform((22050, waveform))
|
76 |
+
|
77 |
+
## GUI space
|
78 |
+
|
79 |
+
title = """
|
80 |
+
<div style="text-align: center; max-width: 700px; margin: 0 auto;">
|
81 |
+
<div
|
82 |
+
style="display: inline-flex; align-items: center; gap: 0.8rem; font-size: 1.75rem;"
|
83 |
+
> <h1 style="font-weight: 900; margin-bottom: 7px; line-height: normal;">
|
84 |
+
VITS2 TTS Catalan Demo
|
85 |
+
</h1> </div>
|
86 |
+
</div>
|
87 |
+
"""
|
88 |
+
|
89 |
+
description = """
|
90 |
+
VITS2 is an end-to-end speech synthesis model that predicts a speech waveform conditional on an input text sequence. VITS2 improved the
|
91 |
+
training and inference efficiency and naturalness by introducing adversarial learning into the duration predictor. The transformer
|
92 |
+
block was added to the normalizing flows to capture the long-term dependency when transforming the distribution.
|
93 |
+
The synthesis quality was improved by incorporating Gaussian noise into the alignment search.
|
94 |
+
|
95 |
+
This model is being trained in openslr69 and festcat datasets
|
96 |
+
"""
|
97 |
+
|
98 |
+
article = "Model by Jungil Kong, et al. from SK telecom. Demo by BSC."
|
99 |
+
|
100 |
+
vits2_inference = gr.Interface(
|
101 |
+
fn=tts,
|
102 |
+
inputs=[
|
103 |
+
gr.Textbox(
|
104 |
+
value="m'ha costat desenvolupar molt una veu, i ara que la tinc no estaré en silenci.",
|
105 |
+
max_lines=1,
|
106 |
+
label="Input text",
|
107 |
+
),
|
108 |
+
gr.Slider(
|
109 |
+
1,
|
110 |
+
47,
|
111 |
+
value=10,
|
112 |
+
step=1,
|
113 |
+
label="Speaker id",
|
114 |
+
info=f"This model is trained on 47 speakers. You can prompt the model using one of these speaker ids.",
|
115 |
+
),
|
116 |
+
gr.Slider(
|
117 |
+
0.5,
|
118 |
+
1.5,
|
119 |
+
value=1,
|
120 |
+
step=0.1,
|
121 |
+
label="Speed",
|
122 |
+
),
|
123 |
+
gr.Slider(
|
124 |
+
0.2,
|
125 |
+
2.0,
|
126 |
+
value=0.667,
|
127 |
+
step=0.01,
|
128 |
+
label="Noise scale",
|
129 |
+
),
|
130 |
+
gr.Slider(
|
131 |
+
0.2,
|
132 |
+
2.0,
|
133 |
+
value=0.8,
|
134 |
+
step=0.01,
|
135 |
+
label="Noise scale w",
|
136 |
+
),
|
137 |
+
],
|
138 |
+
outputs=gr.Audio(),
|
139 |
+
)
|
140 |
+
|
141 |
+
demo = gr.Blocks()
|
142 |
+
|
143 |
+
with demo:
|
144 |
+
gr.Markdown(title)
|
145 |
+
gr.Markdown(description)
|
146 |
+
gr.TabbedInterface([vits2_inference], ["Multispeaker"])
|
147 |
+
gr.Markdown(article)
|
148 |
+
|
149 |
+
demo.queue(max_size=10)
|
150 |
+
demo.launch(show_api=False, server_name="0.0.0.0", server_port=7860)
|