Spaces:
Build error
Build error
Update app.py
Browse files
app.py
CHANGED
@@ -1,17 +1,12 @@
|
|
1 |
import gradio as gr
|
2 |
-
import
|
3 |
-
import random
|
4 |
-
import IPython.display as ipd
|
5 |
import commons
|
6 |
import utils
|
7 |
-
import json
|
8 |
-
import torch
|
9 |
-
import tempfile
|
10 |
-
import numpy as np
|
11 |
from models import SynthesizerTrn
|
12 |
from text.symbols import symbols
|
13 |
from text import text_to_sequence
|
14 |
-
|
|
|
15 |
|
16 |
def get_text(text, hps):
|
17 |
text_norm = text_to_sequence(text, hps.data.text_cleaners)
|
@@ -19,16 +14,6 @@ def get_text(text, hps):
|
|
19 |
text_norm = commons.intersperse(text_norm, 0)
|
20 |
text_norm = torch.LongTensor(text_norm)
|
21 |
return text_norm
|
22 |
-
|
23 |
-
def get_text_byroma(text, hps):
|
24 |
-
text_norm = []
|
25 |
-
for i in text:
|
26 |
-
text_norm.append(symbols.index(i))
|
27 |
-
if hps.data.add_blank:
|
28 |
-
text_norm = commons.intersperse(text_norm, 0)
|
29 |
-
text_norm = torch.LongTensor(text_norm)
|
30 |
-
return text_norm
|
31 |
-
|
32 |
hps = utils.get_hparams_from_file("./configs/leo.json")
|
33 |
net_g = SynthesizerTrn(
|
34 |
len(symbols),
|
@@ -37,61 +22,90 @@ net_g = SynthesizerTrn(
|
|
37 |
n_speakers=hps.data.n_speakers,
|
38 |
**hps.model)
|
39 |
_ = net_g.eval()
|
40 |
-
_ = utils.load_checkpoint("logs/leo/G_4000.pth", net_g, None)
|
41 |
|
42 |
-
|
43 |
-
|
44 |
-
emotion_dict =
|
45 |
-
|
46 |
-
|
47 |
-
""
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
|
|
|
|
52 |
with torch.no_grad():
|
53 |
x_tst = stn_tst.unsqueeze(0)
|
54 |
x_tst_lengths = torch.LongTensor([stn_tst.size(0)])
|
55 |
sid = torch.LongTensor([0])
|
56 |
-
if
|
57 |
-
emo = torch.FloatTensor(
|
|
|
|
|
58 |
elif emotion == "random_sample":
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
break
|
63 |
-
emo = torch.FloatTensor(np.load(f"{random_emotion_root}/{rand_wav}.emo.npy")).unsqueeze(0)
|
64 |
-
print(f"{random_emotion_root}/{rand_wav}")
|
65 |
elif emotion.endswith("wav"):
|
66 |
import emotion_extract
|
67 |
emo = torch.FloatTensor(emotion_extract.extract_wav(emotion))
|
68 |
else:
|
69 |
-
|
|
|
|
|
|
|
|
|
70 |
|
71 |
-
|
|
|
|
|
|
|
|
|
72 |
|
73 |
-
|
74 |
-
|
|
|
|
|
75 |
|
76 |
-
|
77 |
-
ipd.display(ipd.Audio(temp_file_path, rate=hps.data.sampling_rate, normalize=False))
|
78 |
|
79 |
-
|
80 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
-
|
83 |
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
|
91 |
-
|
92 |
-
|
93 |
-
output_audio = gr.outputs.Audio(type="numpy", label="合成音频")
|
94 |
|
95 |
-
|
96 |
-
iface.launch()
|
97 |
|
|
|
1 |
import gradio as gr
|
2 |
+
import torch
|
|
|
|
|
3 |
import commons
|
4 |
import utils
|
|
|
|
|
|
|
|
|
5 |
from models import SynthesizerTrn
|
6 |
from text.symbols import symbols
|
7 |
from text import text_to_sequence
|
8 |
+
import numpy as np
|
9 |
+
|
10 |
|
11 |
def get_text(text, hps):
|
12 |
text_norm = text_to_sequence(text, hps.data.text_cleaners)
|
|
|
14 |
text_norm = commons.intersperse(text_norm, 0)
|
15 |
text_norm = torch.LongTensor(text_norm)
|
16 |
return text_norm
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
hps = utils.get_hparams_from_file("./configs/leo.json")
|
18 |
net_g = SynthesizerTrn(
|
19 |
len(symbols),
|
|
|
22 |
n_speakers=hps.data.n_speakers,
|
23 |
**hps.model)
|
24 |
_ = net_g.eval()
|
|
|
25 |
|
26 |
+
_ = utils.load_checkpoint("./logs/leo/G_4000.pth", net_g, None)
|
27 |
+
all_emotions = np.load("all_emotions.npy")
|
28 |
+
emotion_dict = {
|
29 |
+
"小声": 0,
|
30 |
+
"激动": 1,
|
31 |
+
"平静1": 2,
|
32 |
+
"平静2": 3
|
33 |
+
}
|
34 |
+
import random
|
35 |
+
def tts(txt, emotion):
|
36 |
+
stn_tst = get_text(txt, hps)
|
37 |
+
randsample = None
|
38 |
with torch.no_grad():
|
39 |
x_tst = stn_tst.unsqueeze(0)
|
40 |
x_tst_lengths = torch.LongTensor([stn_tst.size(0)])
|
41 |
sid = torch.LongTensor([0])
|
42 |
+
if type(emotion) ==int:
|
43 |
+
emo = torch.FloatTensor(all_emotions[emotion]).unsqueeze(0)
|
44 |
+
elif emotion == "random":
|
45 |
+
emo = torch.randn([1,1024])
|
46 |
elif emotion == "random_sample":
|
47 |
+
randint = random.randint(0, all_emotions.shape[0])
|
48 |
+
emo = torch.FloatTensor(all_emotions[randint]).unsqueeze(0)
|
49 |
+
randsample = randint
|
|
|
|
|
|
|
50 |
elif emotion.endswith("wav"):
|
51 |
import emotion_extract
|
52 |
emo = torch.FloatTensor(emotion_extract.extract_wav(emotion))
|
53 |
else:
|
54 |
+
emo = torch.FloatTensor(all_emotions[emotion_dict[emotion]]).unsqueeze(0)
|
55 |
+
|
56 |
+
audio = net_g.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=0.667, noise_scale_w=0.8, length_scale=1, emo=emo)[0][0,0].data.float().numpy()
|
57 |
+
return audio, randsample
|
58 |
+
|
59 |
|
60 |
+
def tts1(text, emotion):
|
61 |
+
if len(text) > 150:
|
62 |
+
return "Error: Text is too long", None
|
63 |
+
audio, _ = tts(text, emotion)
|
64 |
+
return "Success", (hps.data.sampling_rate, audio)
|
65 |
|
66 |
+
def tts2(text):
|
67 |
+
if len(text) > 150:
|
68 |
+
return "Error: Text is too long", None
|
69 |
+
audio, randsample = tts(text, "random_sample")
|
70 |
|
71 |
+
return str(randsample), (hps.data.sampling_rate, audio)
|
|
|
72 |
|
73 |
+
def tts3(text, sample):
|
74 |
+
if len(text) > 150:
|
75 |
+
return "Error: Text is too long", None
|
76 |
+
try:
|
77 |
+
audio, _ = tts(text, int(sample))
|
78 |
+
return "Success", (hps.data.sampling_rate, audio)
|
79 |
+
except:
|
80 |
+
return "输入参数不为整数或其他错误", None
|
81 |
+
app = gr.Blocks()
|
82 |
+
with app:
|
83 |
+
with gr.Tabs():
|
84 |
+
with gr.TabItem("使用预制情感合成"):
|
85 |
+
tts_input1 = gr.TextArea(label="日语文本", value="こんにちは。私わあやちねねです。")
|
86 |
+
tts_input2 = gr.Dropdown(label="情感", choices=list(emotion_dict.keys()), value="平静1")
|
87 |
+
tts_submit = gr.Button("合成音频", variant="primary")
|
88 |
+
tts_output1 = gr.Textbox(label="Message")
|
89 |
+
tts_output2 = gr.Audio(label="Output")
|
90 |
+
tts_submit.click(tts1, [tts_input1, tts_input2], [tts_output1, tts_output2])
|
91 |
+
with gr.TabItem("随机抽取训练集样本作为情感参数"):
|
92 |
+
tts_input1 = gr.TextArea(label="日语文本", value="こんにちは。私わあやちねねです。")
|
93 |
+
tts_submit = gr.Button("合成音频", variant="primary")
|
94 |
+
tts_output1 = gr.Textbox(label="随机样本id(可用于第三个tab中合成)")
|
95 |
+
tts_output2 = gr.Audio(label="Output")
|
96 |
+
tts_submit.click(tts2, [tts_input1], [tts_output1, tts_output2])
|
97 |
|
98 |
+
with gr.TabItem("使用情感样本id作为情感参数"):
|
99 |
|
100 |
+
tts_input1 = gr.TextArea(label="日语文本", value="こんにちは。私わあやちねねです。")
|
101 |
+
tts_input2 = gr.Number(label="情感样本id", value=2004)
|
102 |
+
tts_submit = gr.Button("合成音频", variant="primary")
|
103 |
+
tts_output1 = gr.Textbox(label="Message")
|
104 |
+
tts_output2 = gr.Audio(label="Output")
|
105 |
+
tts_submit.click(tts3, [tts_input1, tts_input2], [tts_output1, tts_output2])
|
106 |
|
107 |
+
with gr.TabItem("使用参考音频作为情感参数"):
|
108 |
+
tts_input1 = gr.TextArea(label="text", value="暂未实现")
|
|
|
109 |
|
110 |
+
app.launch()
|
|
|
111 |
|