sayashi commited on
Commit
89b7aca
1 Parent(s): d63c9f2

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +124 -123
app.py CHANGED
@@ -1,123 +1,124 @@
1
- # coding=utf-8
2
- import time
3
- import gradio as gr
4
- import utils
5
- import commons
6
- from models import SynthesizerTrn
7
- from text import text_to_sequence
8
- from torch import no_grad, LongTensor
9
-
10
- hps_ms = utils.get_hparams_from_file(r'./model/config.json')
11
- net_g_ms = SynthesizerTrn(
12
- len(hps_ms.symbols),
13
- hps_ms.data.filter_length // 2 + 1,
14
- hps_ms.train.segment_size // hps_ms.data.hop_length,
15
- n_speakers=hps_ms.data.n_speakers,
16
- **hps_ms.model)
17
- _ = net_g_ms.eval()
18
- speakers = hps_ms.speakers
19
- model, optimizer, learning_rate, epochs = utils.load_checkpoint(r'./model/G_953000.pth', net_g_ms, None)
20
-
21
- def get_text(text, hps):
22
- text_norm, clean_text = text_to_sequence(text, hps.symbols, hps.data.text_cleaners)
23
- if hps.data.add_blank:
24
- text_norm = commons.intersperse(text_norm, 0)
25
- text_norm = LongTensor(text_norm)
26
- return text_norm, clean_text
27
-
28
- def vits(text, language, speaker_id, noise_scale, noise_scale_w, length_scale):
29
- start = time.perf_counter()
30
- if not len(text):
31
- return "输入文本不能为空!", None, None
32
- text = text.replace('\n', ' ').replace('\r', '').replace(" ", "")
33
- if len(text) > 100:
34
- return f"输入文字过长!{len(text)}>100", None, None
35
- if language == 0:
36
- text = f"[ZH]{text}[ZH]"
37
- elif language == 1:
38
- text = f"[JA]{text}[JA]"
39
- else:
40
- text = f"{text}"
41
- stn_tst, clean_text = get_text(text, hps_ms)
42
- with no_grad():
43
- x_tst = stn_tst.unsqueeze(0)
44
- x_tst_lengths = LongTensor([stn_tst.size(0)])
45
- speaker_id = LongTensor([speaker_id])
46
- audio = net_g_ms.infer(x_tst, x_tst_lengths, sid=speaker_id, noise_scale=noise_scale, noise_scale_w=noise_scale_w,
47
- length_scale=length_scale)[0][0, 0].data.float().numpy()
48
-
49
- return "生成成功!", (22050, audio), f"生成耗时 {round(time.perf_counter()-start, 2)} s"
50
-
51
- def search_speaker(search_value):
52
- for s in speakers:
53
- if search_value == s:
54
- return s
55
- for s in speakers:
56
- if search_value in s:
57
- return s
58
-
59
- def change_lang(language):
60
- if language == 0:
61
- return 0.6, 0.668, 1.2
62
- else:
63
- return 0.6, 0.668, 1.1
64
-
65
- download_audio_js = """
66
- () =>{{
67
- let root = document.querySelector("body > gradio-app");
68
- if (root.shadowRoot != null)
69
- root = root.shadowRoot;
70
- let audio = root.querySelector("#tts-audio").querySelector("audio");
71
- let text = root.querySelector("#input-text").querySelector("textarea");
72
- if (audio == undefined)
73
- return;
74
- text = text.value;
75
- if (text == undefined)
76
- text = Math.floor(Math.random()*100000000);
77
- audio = audio.src;
78
- let oA = document.createElement("a");
79
- oA.download = text.substr(0, 20)+'.wav';
80
- oA.href = audio;
81
- document.body.appendChild(oA);
82
- oA.click();
83
- oA.remove();
84
- }}
85
- """
86
-
87
- if __name__ == '__main__':
88
- with gr.Blocks() as app:
89
- gr.Markdown(
90
- "# <center> VITS语音在线合成demo\n"
91
- "<div align='center'>主要有赛马娘,原神中文,原神日语,崩坏3的音色</div>"
92
- '<div align="center"><a><font color="#dd0000">结果有随机性,语调可能很奇怪,可多次生成取最佳效果</font></a></div>'
93
- '<div align="center"><a><font color="#dd0000">标点符号会影响生成的结果</font></a></div>'
94
- )
95
-
96
- with gr.Tabs():
97
- with gr.TabItem("vits"):
98
- with gr.Row():
99
- with gr.Column():
100
- input_text = gr.Textbox(label="Text (100 words limitation)", lines=5, value="今天晚上吃啥好呢。", elem_id=f"input-text")
101
- lang = gr.Dropdown(label="Language", choices=["中文", "日语", "中日混合(中文用[ZH][ZH]包裹起来,日文用[JA][JA]包裹起来)"],
102
- type="index", value="中文")
103
- btn = gr.Button(value="Submit")
104
- with gr.Row():
105
- search = gr.Textbox(label="Search Speaker", lines=1)
106
- btn2 = gr.Button(value="Search")
107
- sid = gr.Dropdown(label="Speaker", choices=speakers, type="index", value=speakers[228])
108
- with gr.Row():
109
- ns = gr.Slider(label="noise_scale(控制感情变化程度)", minimum=0.1, maximum=1.0, step=0.1, value=0.6, interactive=True)
110
- nsw = gr.Slider(label="noise_scale_w(控制音素发音长度)", minimum=0.1, maximum=1.0, step=0.1, value=0.668, interactive=True)
111
- ls = gr.Slider(label="length_scale(控制整体语速)", minimum=0.1, maximum=2.0, step=0.1, value=1.2, interactive=True)
112
- with gr.Column():
113
- o1 = gr.Textbox(label="Output Message")
114
- o2 = gr.Audio(label="Output Audio", elem_id=f"tts-audio")
115
- o3 = gr.Textbox(label="Extra Info")
116
- download = gr.Button("Download Audio")
117
- btn.click(vits, inputs=[input_text, lang, sid, ns, nsw, ls], outputs=[o1, o2, o3])
118
- download.click(None, [], [], _js=download_audio_js.format())
119
- btn2.click(search_speaker, inputs=[search], outputs=[sid])
120
- lang.change(change_lang, inputs=[lang], outputs=[ns, nsw, ls])
121
- with gr.TabItem("可用人物一览"):
122
- gr.Radio(label="Speaker", choices=speakers, interactive=False, type="index")
123
- app.queue(concurrency_count=1).launch()
 
 
1
+ # coding=utf-8
2
+ import time
3
+ import gradio as gr
4
+ import utils
5
+ import commons
6
+ from models import SynthesizerTrn
7
+ from text import text_to_sequence
8
+ from torch import no_grad, LongTensor
9
+
10
+ hps_ms = utils.get_hparams_from_file(r'./model/config.json')
11
+ net_g_ms = SynthesizerTrn(
12
+ len(hps_ms.symbols),
13
+ hps_ms.data.filter_length // 2 + 1,
14
+ hps_ms.train.segment_size // hps_ms.data.hop_length,
15
+ n_speakers=hps_ms.data.n_speakers,
16
+ **hps_ms.model)
17
+ _ = net_g_ms.eval()
18
+ speakers = hps_ms.speakers
19
+ model, optimizer, learning_rate, epochs = utils.load_checkpoint(r'./model/G_953000.pth', net_g_ms, None)
20
+
21
+ def get_text(text, hps):
22
+ text_norm, clean_text = text_to_sequence(text, hps.symbols, hps.data.text_cleaners)
23
+ if hps.data.add_blank:
24
+ text_norm = commons.intersperse(text_norm, 0)
25
+ text_norm = LongTensor(text_norm)
26
+ return text_norm, clean_text
27
+
28
+ def vits(text, language, speaker_id, noise_scale, noise_scale_w, length_scale):
29
+ start = time.perf_counter()
30
+ if not len(text):
31
+ return "输入文本不能为空!", None, None
32
+ text = text.replace('\n', ' ').replace('\r', '').replace(" ", "")
33
+ if len(text) > 100:
34
+ return f"输入文字过长!{len(text)}>100", None, None
35
+ if language == 0:
36
+ text = f"[ZH]{text}[ZH]"
37
+ elif language == 1:
38
+ text = f"[JA]{text}[JA]"
39
+ else:
40
+ text = f"{text}"
41
+ stn_tst, clean_text = get_text(text, hps_ms)
42
+ with no_grad():
43
+ x_tst = stn_tst.unsqueeze(0)
44
+ x_tst_lengths = LongTensor([stn_tst.size(0)])
45
+ speaker_id = LongTensor([speaker_id])
46
+ audio = net_g_ms.infer(x_tst, x_tst_lengths, sid=speaker_id, noise_scale=noise_scale, noise_scale_w=noise_scale_w,
47
+ length_scale=length_scale)[0][0, 0].data.float().numpy()
48
+
49
+ return "生成成功!", (22050, audio), f"生成耗时 {round(time.perf_counter()-start, 2)} s"
50
+
51
+ def search_speaker(search_value):
52
+ for s in speakers:
53
+ if search_value == s:
54
+ return s
55
+ for s in speakers:
56
+ if search_value in s:
57
+ return s
58
+
59
+ def change_lang(language):
60
+ if language == 0:
61
+ return 0.6, 0.668, 1.2
62
+ else:
63
+ return 0.6, 0.668, 1.1
64
+
65
+ download_audio_js = """
66
+ () =>{{
67
+ let root = document.querySelector("body > gradio-app");
68
+ if (root.shadowRoot != null)
69
+ root = root.shadowRoot;
70
+ let audio = root.querySelector("#tts-audio").querySelector("audio");
71
+ let text = root.querySelector("#input-text").querySelector("textarea");
72
+ if (audio == undefined)
73
+ return;
74
+ text = text.value;
75
+ if (text == undefined)
76
+ text = Math.floor(Math.random()*100000000);
77
+ audio = audio.src;
78
+ let oA = document.createElement("a");
79
+ oA.download = text.substr(0, 20)+'.wav';
80
+ oA.href = audio;
81
+ document.body.appendChild(oA);
82
+ oA.click();
83
+ oA.remove();
84
+ }}
85
+ """
86
+
87
+ if __name__ == '__main__':
88
+ with gr.Blocks() as app:
89
+ gr.Markdown(
90
+ "# <center> VITS语音在线合成demo\n"
91
+ "# <center> 严禁将模型用于任何商业项目,否则后果自负\n"
92
+ "<div align='center'>主要有赛马娘,原神中文,原神日语,崩坏3的音色</div>"
93
+ '<div align="center"><a><font color="#dd0000">结果有随机性,语调可能很奇怪,可多次生成取最佳效果</font></a></div>'
94
+ '<div align="center"><a><font color="#dd0000">标点符号会影响生成的结果</font></a></div>'
95
+ )
96
+
97
+ with gr.Tabs():
98
+ with gr.TabItem("vits"):
99
+ with gr.Row():
100
+ with gr.Column():
101
+ input_text = gr.Textbox(label="Text (100 words limitation)", lines=5, value="今天晚上吃啥好呢。", elem_id=f"input-text")
102
+ lang = gr.Dropdown(label="Language", choices=["中文", "日语", "中日混合(中文用[ZH][ZH]包裹起来,日文用[JA][JA]包裹起来)"],
103
+ type="index", value="中文")
104
+ btn = gr.Button(value="Submit")
105
+ with gr.Row():
106
+ search = gr.Textbox(label="Search Speaker", lines=1)
107
+ btn2 = gr.Button(value="Search")
108
+ sid = gr.Dropdown(label="Speaker", choices=speakers, type="index", value=speakers[228])
109
+ with gr.Row():
110
+ ns = gr.Slider(label="noise_scale(控制感情变化程度)", minimum=0.1, maximum=1.0, step=0.1, value=0.6, interactive=True)
111
+ nsw = gr.Slider(label="noise_scale_w(控制音素发音长度)", minimum=0.1, maximum=1.0, step=0.1, value=0.668, interactive=True)
112
+ ls = gr.Slider(label="length_scale(控制整体语速)", minimum=0.1, maximum=2.0, step=0.1, value=1.2, interactive=True)
113
+ with gr.Column():
114
+ o1 = gr.Textbox(label="Output Message")
115
+ o2 = gr.Audio(label="Output Audio", elem_id=f"tts-audio")
116
+ o3 = gr.Textbox(label="Extra Info")
117
+ download = gr.Button("Download Audio")
118
+ btn.click(vits, inputs=[input_text, lang, sid, ns, nsw, ls], outputs=[o1, o2, o3])
119
+ download.click(None, [], [], _js=download_audio_js.format())
120
+ btn2.click(search_speaker, inputs=[search], outputs=[sid])
121
+ lang.change(change_lang, inputs=[lang], outputs=[ns, nsw, ls])
122
+ with gr.TabItem("可用人物一览"):
123
+ gr.Radio(label="Speaker", choices=speakers, interactive=False, type="index")
124
+ app.queue(concurrency_count=1).launch()