File size: 9,747 Bytes
773c7bd
 
376b5d9
773c7bd
376b5d9
 
773c7bd
 
 
f784787
376b5d9
 
 
773c7bd
 
 
 
376b5d9
f784787
773c7bd
376b5d9
773c7bd
 
 
376b5d9
773c7bd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f58d262
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f784787
773c7bd
 
 
 
f58d262
773c7bd
c837795
f58d262
c837795
 
773c7bd
f58d262
773c7bd
c837795
773c7bd
c837795
f58d262
 
c837795
 
 
773c7bd
 
c837795
773c7bd
 
c837795
 
 
 
 
773c7bd
c837795
 
 
f58d262
c837795
773c7bd
f58d262
c837795
 
 
f58d262
 
 
c837795
 
 
 
 
 
 
 
 
f58d262
c837795
 
 
 
 
 
 
 
 
f58d262
 
 
 
 
c837795
 
 
 
 
 
 
 
773c7bd
c837795
 
 
 
 
 
773c7bd
 
c837795
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
773c7bd
 
c837795
 
 
 
 
 
 
 
 
773c7bd
 
c837795
 
 
 
773c7bd
c837795
 
 
 
 
f58d262
 
773c7bd
c837795
 
376b5d9
 
 
 
 
773c7bd
 
 
 
 
 
 
7ca618f
 
 
773c7bd
 
 
 
 
 
 
 
 
 
 
f58d262
773c7bd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f58d262
 
 
 
773c7bd
 
 
5d8cb3b
773c7bd
 
 
 
 
 
 
 
 
 
 
 
 
376b5d9
773c7bd
376b5d9
 
773c7bd
 
 
376b5d9
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
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
import csv
import datetime
import os
import re
import time
import uuid
from io import StringIO

import gradio as gr
import spaces
import torch
import torchaudio
from huggingface_hub import HfApi, hf_hub_download, snapshot_download
from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import Xtts
from vinorm import TTSnorm

# download for mecab
os.system("python -m unidic download")

HF_TOKEN = os.environ.get("HF_TOKEN")
api = HfApi(token=HF_TOKEN)

# This will trigger downloading model
print("Downloading if not downloaded viXTTS")
checkpoint_dir = "model/"
repo_id = "capleaf/viXTTS"
use_deepspeed = False

os.makedirs(checkpoint_dir, exist_ok=True)

required_files = ["model.pth", "config.json", "vocab.json", "speakers_xtts.pth"]
files_in_dir = os.listdir(checkpoint_dir)
if not all(file in files_in_dir for file in required_files):
    snapshot_download(
        repo_id=repo_id,
        repo_type="model",
        local_dir=checkpoint_dir,
    )
    hf_hub_download(
        repo_id="coqui/XTTS-v2",
        filename="speakers_xtts.pth",
        local_dir=checkpoint_dir,
    )

xtts_config = os.path.join(checkpoint_dir, "config.json")
config = XttsConfig()
config.load_json(xtts_config)
MODEL = Xtts.init_from_config(config)
MODEL.load_checkpoint(
    config, checkpoint_dir=checkpoint_dir, use_deepspeed=use_deepspeed
)
if torch.cuda.is_available():
    MODEL.cuda()

supported_languages = config.languages
if not "vi" in supported_languages:
    supported_languages.append("vi")


def normalize_vietnamese_text(text):
    text = (
        TTSnorm(text, unknown=False, lower=False, rule=True)
        .replace("..", ".")
        .replace("!.", "!")
        .replace("?.", "?")
        .replace(" .", ".")
        .replace(" ,", ",")
        .replace('"', "")
        .replace("'", "")
        .replace("AI", "Ây Ai")
        .replace("A.I", "Ây Ai")
    )
    return text


def calculate_keep_len(text, lang):
    """Simple hack for short sentences"""
    if lang in ["ja", "zh-cn"]:
        return -1

    word_count = len(text.split())
    num_punct = text.count(".") + text.count("!") + text.count("?") + text.count(",")

    if word_count < 5:
        return 15000 * word_count + 2000 * num_punct
    elif word_count < 10:
        return 13000 * word_count + 2000 * num_punct
    return -1


@spaces.GPU
def predict(
    prompt,
    language,
    audio_file_pth,
    normalize_text=True,
):
    if language not in supported_languages:
        metrics_text = gr.Warning(
            f"Language you put {language} in is not in is not in our Supported Languages, please choose from dropdown"
        )

        return (None, metrics_text)

    speaker_wav = audio_file_pth

    if len(prompt) < 2:
        metrics_text = gr.Warning("Please give a longer prompt text")
        return (None, metrics_text)
    try:
        metrics_text = ""
        t_latent = time.time()

        try:
            (
                gpt_cond_latent,
                speaker_embedding,
            ) = MODEL.get_conditioning_latents(
                audio_path=speaker_wav,
                gpt_cond_len=30,
                gpt_cond_chunk_len=4,
                max_ref_length=60,
            )

        except Exception as e:
            print("Speaker encoding error", str(e))
            metrics_text = gr.Warning(
                "It appears something wrong with reference, did you unmute your microphone?"
            )
            return (None, metrics_text)

        prompt = re.sub("([^\x00-\x7F]|\w)(\.|\。|\?)", r"\1 \2\2", prompt)

        if normalize_text and language == "vi":
            prompt = normalize_vietnamese_text(prompt)

        print("I: Generating new audio...")
        t0 = time.time()
        out = MODEL.inference(
            prompt,
            language,
            gpt_cond_latent,
            speaker_embedding,
            repetition_penalty=5.0,
            temperature=0.75,
            enable_text_splitting=True,
        )
        inference_time = time.time() - t0
        print(f"I: Time to generate audio: {round(inference_time*1000)} milliseconds")
        metrics_text += (
            f"Time to generate audio: {round(inference_time*1000)} milliseconds\n"
        )
        real_time_factor = (time.time() - t0) / out["wav"].shape[-1] * 24000
        print(f"Real-time factor (RTF): {real_time_factor}")
        metrics_text += f"Real-time factor (RTF): {real_time_factor:.2f}\n"

        # Temporary hack for short sentences
        keep_len = calculate_keep_len(prompt, language)
        out["wav"] = out["wav"][:keep_len]

        torchaudio.save("output.wav", torch.tensor(out["wav"]).unsqueeze(0), 24000)

    except RuntimeError as e:
        if "device-side assert" in str(e):
            # cannot do anything on cuda device side error, need tor estart
            print(
                f"Exit due to: Unrecoverable exception caused by language:{language} prompt:{prompt}",
                flush=True,
            )
            gr.Warning("Unhandled Exception encounter, please retry in a minute")
            print("Cuda device-assert Runtime encountered need restart")

            error_time = datetime.datetime.now().strftime("%d-%m-%Y-%H:%M:%S")
            error_data = [
                error_time,
                prompt,
                language,
                audio_file_pth,
            ]
            error_data = [str(e) if type(e) != str else e for e in error_data]
            print(error_data)
            print(speaker_wav)
            write_io = StringIO()
            csv.writer(write_io).writerows([error_data])
            csv_upload = write_io.getvalue().encode()

            filename = error_time + "_" + str(uuid.uuid4()) + ".csv"
            print("Writing error csv")
            error_api = HfApi()
            error_api.upload_file(
                path_or_fileobj=csv_upload,
                path_in_repo=filename,
                repo_id="coqui/xtts-flagged-dataset",
                repo_type="dataset",
            )

            # speaker_wav
            print("Writing error reference audio")
            speaker_filename = error_time + "_reference_" + str(uuid.uuid4()) + ".wav"
            error_api = HfApi()
            error_api.upload_file(
                path_or_fileobj=speaker_wav,
                path_in_repo=speaker_filename,
                repo_id="coqui/xtts-flagged-dataset",
                repo_type="dataset",
            )

            # HF Space specific.. This error is unrecoverable need to restart space
            space = api.get_space_runtime(repo_id=repo_id)
            if space.stage != "BUILDING":
                api.restart_space(repo_id=repo_id)
            else:
                print("TRIED TO RESTART but space is building")

        else:
            if "Failed to decode" in str(e):
                print("Speaker encoding error", str(e))
                metrics_text = gr.Warning(
                    metrics_text="It appears something wrong with reference, did you unmute your microphone?"
                )
            else:
                print("RuntimeError: non device-side assert error:", str(e))
                metrics_text = gr.Warning(
                    "Something unexpected happened please retry again."
                )
            return (None, metrics_text)
    return ("output.wav", metrics_text)


with gr.Blocks(analytics_enabled=False) as demo:
    with gr.Row():
        with gr.Column():
            gr.Markdown(
                """
                # viXTTS Demo ✨
                - Github: https://github.com/thinhlpg/vixtts-demo/
                - viVoice: https://github.com/thinhlpg/viVoice
                """
            )
        with gr.Column():
            # placeholder to align the image
            pass

    with gr.Row():
        with gr.Column():
            input_text_gr = gr.Textbox(
                label="Text Prompt",
                info="One or two sentences at a time is better. Up to 200 text characters.",
                value="Xin chào, tôi là một mô hình chuyển đổi văn bản thành giọng nói tiếng Việt",
            )
            language_gr = gr.Dropdown(
                label="Language",
                info="Select an output language for the synthesised speech",
                choices=[
                    "vi",
                    "en",
                    "es",
                    "fr",
                    "de",
                    "it",
                    "pt",
                    "pl",
                    "tr",
                    "ru",
                    "nl",
                    "cs",
                    "ar",
                    "zh-cn",
                    "ja",
                    "ko",
                    "hu",
                    "hi",
                ],
                max_choices=1,
                value="vi",
            )
            normalize_text = gr.Checkbox(
                label="Normalize Vietnamese Text",
                info="Normalize Vietnamese Text",
            )
            ref_gr = gr.Audio(
                label="Reference Audio",
                type="filepath",
                value="model/samples/nu-luu-loat.wav",
            )
            tts_button = gr.Button("Send", elem_id="send-btn", visible=True)

        with gr.Column():
            audio_gr = gr.Audio(label="Synthesised Audio", autoplay=True)
            out_text_gr = gr.Text(label="Metrics")

    tts_button.click(
        predict,
        [
            input_text_gr,
            language_gr,
            ref_gr,
            normalize_text,
        ],
        outputs=[audio_gr, out_text_gr],
        api_name="predict",
    )

demo.queue()
demo.launch(debug=True, show_api=True, share=True)