File size: 11,352 Bytes
f5915fd
4028449
f5915fd
4028449
 
 
f5915fd
635f007
 
1567c2d
d430de8
a520615
85fcd3c
73dbaa9
a520615
641afe5
635f007
 
1c22955
 
 
e2e4977
4028449
 
50a9d0f
6eb9ea3
 
 
 
 
 
b21342c
 
 
 
 
 
 
 
 
f5915fd
b21342c
e27b102
 
138b27f
 
fbe2075
 
 
85fcd3c
e27b102
 
 
62de84c
e27b102
 
a5cfbc2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e1d1d80
6440f80
 
 
 
 
 
 
e2e4977
 
138b27f
 
efc3af1
 
fbe2075
 
 
3ecf742
6440f80
efc3af1
dc39974
efc3af1
6440f80
efc3af1
 
6440f80
4a6da64
6440f80
 
 
 
 
 
4028449
d430de8
 
138b27f
 
fbe2075
 
 
3ecf742
6440f80
 
4a6da64
6440f80
d430de8
635f007
04c285e
d430de8
 
 
39e26fe
4028449
addff22
d430de8
 
04c285e
4b8ade9
e2e4977
 
 
 
04c285e
cf9af86
23a6d4b
e1d1d80
 
e2e4977
 
04c285e
e1d1d80
a5cfbc2
 
 
 
 
 
 
 
 
 
 
4b8ade9
 
 
 
4a6da64
cf9af86
4b8ade9
 
04c285e
d3f7823
641afe5
f5915fd
635f007
4028449
e7f95df
50a9d0f
5967c32
50a9d0f
 
 
 
c609d03
50a9d0f
769b7d2
635f007
304fce9
836bb59
635f007
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
INTROTXT = """# StyleTTS 2
kudos to mrfakename for the base gradio code I'm borrowing here.

日本語用
The Text-guided inference may or may not work. you can only do inference max 512 tokens.
**
"""
import gradio as gr
import styletts2importable
import ljspeechimportable
import torch
import os
from txtsplit import txtsplit
import numpy as np
import pickle
theme = gr.themes.Base(
    font=[gr.themes.GoogleFont('Libre Franklin'), gr.themes.GoogleFont('Public Sans'), 'system-ui', 'sans-serif'],
)

VO_JA_Kamisato_Ayaka_About_Kujou_Sara.wav
voicelist = ['VO_JA_Kamisato_Ayaka_About_Kujou_Sara','hontonokimochi','gaen_original']
voices = {}
# import phonemizer
# global_phonemizer = phonemizer.backend.EspeakBackend(language='en-us', preserve_punctuation=True,  with_stress=True)
# todo: cache computed style, load using pickle
# if os.path.exists('voices.pkl'):
    # with open('voices.pkl', 'rb') as f:
        # voices = pickle.load(f)
# else:
for v in voicelist:
    voices[v] = styletts2importable.compute_style(f'voices/{v}.wav')
# def synthesize(text, voice, multispeakersteps):
#     if text.strip() == "":
#         raise gr.Error("You must enter some text")
#     # if len(global_phonemizer.phonemize([text])) > 300:
#     if len(text) > 300:
#         raise gr.Error("Text must be under 300 characters")
#     v = voice.lower()
#     # return (24000, styletts2importable.inference(text, voices[v], alpha=0.3, beta=0.7, diffusion_steps=7, embedding_scale=1))
#     return (24000, styletts2importable.inference(text, voices[v], alpha=0.3, beta=0.7, diffusion_steps=multispeakersteps, embedding_scale=1))
if not torch.cuda.is_available(): INTROTXT += "\n\n### You are on a CPU-only system, inference will be much slower.\n\nYou can use the [online demo](https://huggingface.co/spaces/styletts2/styletts2) for fast inference."
def synthesize(text, voice, lngsteps, password, progress=gr.Progress()):
    if text.strip() == "":
        raise gr.Error("You must enter some text")
    if len(text) > 50000:
        raise gr.Error("Text must be <50k characters")
    print("*** saying ***")
    print(text)
    print("*** end ***")
    texts = txtsplit(text)
    v = voice.lower()
    audios = []
    for t in progress.tqdm(texts):
        print(t)
        audios.append(styletts2importable.inference(t, voices[v], alpha=0.3, beta=0.7, diffusion_steps=lngsteps, embedding_scale=1))
    return (24000, np.concatenate(audios))
# def longsynthesize(text, voice, lngsteps, password, progress=gr.Progress()):
#     if password == os.environ['ACCESS_CODE']:
#         if text.strip() == "":
#             raise gr.Error("You must enter some text")
#         if lngsteps > 25:
#             raise gr.Error("Max 25 steps")
#         if lngsteps < 5:
#             raise gr.Error("Min 5 steps")
#         texts = split_and_recombine_text(text)
#         v = voice.lower()
#         audios = []
#         for t in progress.tqdm(texts):
#             audios.append(styletts2importable.inference(t, voices[v], alpha=0.3, beta=0.7, diffusion_steps=lngsteps, embedding_scale=1))
#         return (24000, np.concatenate(audios))
#     else:
#         raise gr.Error('Wrong access code')
def clsynthesize(text, voice, vcsteps, embscale, alpha, beta, progress=gr.Progress()):
    # if text.strip() == "":
    #     raise gr.Error("You must enter some text")
    # # if global_phonemizer.phonemize([text]) > 300:
    # if len(text) > 400:
    #     raise gr.Error("Text must be under 400 characters")
    # # return (24000, styletts2importable.inference(text, styletts2importable.compute_style(voice), alpha=0.3, beta=0.7, diffusion_steps=20, embedding_scale=1))
    # return (24000, styletts2importable.inference(text, styletts2importable.compute_style(voice), alpha=0.3, beta=0.7, diffusion_steps=vcsteps, embedding_scale=1))
    if text.strip() == "":
        raise gr.Error("You must enter some text")
    if len(text) > 50000:
        raise gr.Error("Text must be <50k characters")
    if embscale > 1.3 and len(text) < 20:
        gr.Warning("WARNING: You entered short text, you may get static!")
    print("*** saying ***")
    print(text)
    print("*** end ***")
    texts = txtsplit(text)
    audios = []
    # vs = styletts2importable.compute_style(voice)
    vs = styletts2importable.compute_style(voice)
    # print(vs)
    for t in progress.tqdm(texts):
        audios.append(styletts2importable.inference(t, vs, alpha=alpha, beta=beta, diffusion_steps=vcsteps, embedding_scale=embscale))
        # audios.append(styletts2importable.inference(t, vs, diffusion_steps=10, alpha=0.3, beta=0.7, embedding_scale=5))
    return (24000, np.concatenate(audios))
def ljsynthesize(text, steps,embscale, progress=gr.Progress()):
    # if text.strip() == "":
    #     raise gr.Error("You must enter some text")
    # # if global_phonemizer.phonemize([text]) > 300:
    # if len(text) > 400:
    #     raise gr.Error("Text must be under 400 characters")
    noise = torch.randn(1,1,256).to('cuda' if torch.cuda.is_available() else 'cpu')
    # return (24000, Text-guided Inferenceimportable.inference(text, noise, diffusion_steps=7, embedding_scale=1))
    if text.strip() == "":
        raise gr.Error("You must enter some text")
    if len(text) > 150000:
        raise gr.Error("Text must be <150k characters")
    print("*** saying ***")
    print(text)
    print("*** end ***")
    texts = txtsplit(text)
    audios = []
    for t in progress.tqdm(texts):
        audios.append(ljspeechimportable.inference(t, noise, diffusion_steps=steps, embedding_scale=embscale))
    return (24000, np.concatenate(audios))


with gr.Blocks() as vctk:
    with gr.Row():
        with gr.Column(scale=1):
            inp = gr.Textbox(label="Text", info="What would you like StyleTTS 2 to read? It works better on full sentences.", interactive=True)
            voice = gr.Dropdown(voicelist, label="Voice", info="Select a default voice.", value='m-us-2', interactive=True)
            multispeakersteps = gr.Slider(minimum=3, maximum=15, value=3, step=1, label="Diffusion Steps", info="Higher gives you more diverse results but not necessarily higher quality - これを増えたらもっとエモーショナルな結果になりますが、クオリティーのいい結果になるとは限らない。", interactive=True)
            # use_gruut = gr.Checkbox(label="Use alternate phonemizer (Gruut) - Experimental")
        with gr.Column(scale=1):
            btn = gr.Button("Synthesize", variant="primary")
            audio = gr.Audio(interactive=False, label="Synthesized Audio", waveform_options={'waveform_progress_color': '#3C82F6'})
            btn.click(synthesize, inputs=[inp, voice, multispeakersteps], outputs=[audio], concurrency_limit=4)
with gr.Blocks() as clone:
    with gr.Row():
        with gr.Column(scale=1):
            clinp = gr.Textbox(label="Text", info="What would you like StyleTTS 2 to read? It works better on full sentences.", interactive=True)
            clvoice = gr.Audio(label="Voice", interactive=True, type='filepath', max_length=300, waveform_options={'waveform_progress_color': '#3C82F6'})
            vcsteps = gr.Slider(minimum=3, maximum=20, value=20, step=1, label="Diffusion Steps", info="Theoretically, higher should be better quality but slower, but we cannot notice a difference. Try with lower steps first - it is faster", interactive=True)
            embscale = gr.Slider(minimum=1, maximum=10, value=1, step=0.1, label="Embedding Scale (READ WARNING BELOW)", info="Defaults to 1. WARNING: If you set this too high and generate text that's too short you will get static!", interactive=True)
            alpha = gr.Slider(minimum=0, maximum=1, value=0.3, step=0.1, label="Alpha", info="Defaults to 0.3", interactive=True)
            beta = gr.Slider(minimum=0, maximum=1, value=0.7, step=0.1, label="Beta", info="Defaults to 0.7", interactive=True)
        with gr.Column(scale=1):
            clbtn = gr.Button("Synthesize", variant="primary")
            claudio = gr.Audio(interactive=False, label="Synthesized Audio", waveform_options={'waveform_progress_color': '#3C82F6'})
            clbtn.click(clsynthesize, inputs=[clinp, clvoice, vcsteps, embscale, alpha, beta], outputs=[claudio], concurrency_limit=4)
# with gr.Blocks() as longText:
#     with gr.Row():
#         with gr.Column(scale=1):
#             lnginp = gr.Textbox(label="Text", info="What would you like StyleTTS 2 to read? It works better on full sentences.", interactive=True)
#             lngvoice = gr.Dropdown(voicelist, label="Voice", info="Select a default voice.", value='m-us-1', interactive=True)
#             lngsteps = gr.Slider(minimum=5, maximum=25, value=10, step=1, label="Diffusion Steps", info="Higher = better quality, but slower", interactive=True)
#             lngpwd = gr.Textbox(label="Access code", info="This feature is in beta. You need an access code to use it as it uses more resources and we would like to prevent abuse")
#         with gr.Column(scale=1):
#             lngbtn = gr.Button("Synthesize", variant="primary")
#             lngaudio = gr.Audio(interactive=False, label="Synthesized Audio")
#             lngbtn.click(longsynthesize, inputs=[lnginp, lngvoice, lngsteps, lngpwd], outputs=[lngaudio], concurrency_limit=4)
with gr.Blocks() as lj:
    with gr.Row():
        with gr.Column(scale=1):
            ljinp = gr.Textbox(label="Text", info="What would you like StyleTTS 2 to read? It works better on full sentences.", interactive=True)
            embscale = gr.Slider(minimum=1, maximum=10, value=1, step=0.1, label="Embedding Scale (READ WARNING BELOW)", info="Defaults to 1. これを上げたらパフォーマンスがもっとエモーショナルになる、増やしすぎるとだめになるので、ご注意ください", interactive=True)
            ljsteps = gr.Slider(minimum=3, maximum=20, value=3, step=1, label="Diffusion Steps", info="Theoretically, higher should be better quality but slower, but we cannot notice a difference. Try with lower steps first - it is faster", interactive=True)
        with gr.Column(scale=1):
            ljbtn = gr.Button("Synthesize", variant="primary")
            ljaudio = gr.Audio(interactive=False, label="Synthesized Audio", waveform_options={'waveform_progress_color': '#3C82F6'})
            ljbtn.click(ljsynthesize, inputs=[ljinp, ljsteps, embscale], outputs=[ljaudio], concurrency_limit=4)
with gr.Blocks(title="StyleTTS 2", css="footer{display:none !important}", theme="NoCrypt/miku") as demo:
    gr.Markdown(INTROTXT)
    gr.DuplicateButton("Duplicate Space")
    # gr.TabbedInterface([vctk, clone, lj, longText], ['Multi-Voice', 'Voice Cloning', 'Text-guided Inference', 'Long Text [Beta]'])
    gr.TabbedInterface([vctk, clone, lj], ['Multi-Voice', 'do not use this option','Text-guided Inference', 'Long Text [Beta]'])
    gr.Markdown("""
Demo by [mrfakename](https://twitter.com/realmrfakename). I am not affiliated with the StyleTTS 2 authors.

Run this demo locally using Docker:

```bash
docker run -it -p 7860:7860 --platform=linux/amd64 --gpus all registry.hf.space/styletts2-styletts2:latest python app.py
```
""") # Please do not remove this line.
if __name__ == "__main__":
    # demo.queue(api_open=False, max_size=15).launch(show_api=False)
    demo.queue(api_open=False, max_size=15).launch(show_api=False)