Update app.py
Browse files
app.py
CHANGED
@@ -14,14 +14,12 @@ from TTS.tts.configs.xtts_config import XttsConfig
|
|
14 |
from TTS.tts.models.xtts import Xtts
|
15 |
from vinorm import TTSnorm
|
16 |
|
17 |
-
#
|
18 |
-
# os.system("python -m unidic download")
|
19 |
-
|
20 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
21 |
api = HfApi(token=HF_TOKEN)
|
22 |
|
23 |
-
#
|
24 |
-
print("Downloading if not
|
25 |
checkpoint_dir = "model/"
|
26 |
repo_id = "capleaf/viXTTS"
|
27 |
use_deepspeed = False
|
@@ -42,6 +40,7 @@ if not all(file in files_in_dir for file in required_files):
|
|
42 |
local_dir=checkpoint_dir,
|
43 |
)
|
44 |
|
|
|
45 |
xtts_config = os.path.join(checkpoint_dir, "config.json")
|
46 |
config = XttsConfig()
|
47 |
config.load_json(xtts_config)
|
@@ -52,8 +51,9 @@ MODEL.load_checkpoint(
|
|
52 |
if torch.cuda.is_available():
|
53 |
MODEL.cuda()
|
54 |
|
|
|
55 |
supported_languages = config.languages
|
56 |
-
if
|
57 |
supported_languages.append("vi")
|
58 |
|
59 |
|
@@ -74,7 +74,6 @@ def normalize_vietnamese_text(text):
|
|
74 |
|
75 |
|
76 |
def calculate_keep_len(text, lang):
|
77 |
-
"""Simple hack for short sentences"""
|
78 |
if lang in ["ja", "zh-cn"]:
|
79 |
return -1
|
80 |
|
@@ -88,52 +87,39 @@ def calculate_keep_len(text, lang):
|
|
88 |
return -1
|
89 |
|
90 |
|
91 |
-
def predict(
|
92 |
-
prompt,
|
93 |
-
language,
|
94 |
-
audio_file_pth,
|
95 |
-
normalize_text=True,
|
96 |
-
):
|
97 |
if language not in supported_languages:
|
98 |
metrics_text = gr.Warning(
|
99 |
-
f"Language
|
100 |
)
|
101 |
-
|
102 |
-
return (None, metrics_text)
|
103 |
-
|
104 |
-
speaker_wav = audio_file_pth
|
105 |
|
106 |
if len(prompt) < 2:
|
107 |
-
metrics_text = gr.Warning("Please
|
108 |
-
return
|
|
|
109 |
try:
|
110 |
metrics_text = ""
|
111 |
t_latent = time.time()
|
112 |
|
113 |
try:
|
114 |
-
(
|
115 |
-
|
116 |
-
speaker_embedding,
|
117 |
-
) = MODEL.get_conditioning_latents(
|
118 |
-
audio_path=speaker_wav,
|
119 |
gpt_cond_len=30,
|
120 |
gpt_cond_chunk_len=4,
|
121 |
max_ref_length=60,
|
122 |
)
|
123 |
-
|
124 |
except Exception as e:
|
125 |
-
print("Speaker encoding error", str(e))
|
126 |
-
metrics_text = gr.Warning(
|
127 |
-
|
128 |
-
)
|
129 |
-
return (None, metrics_text)
|
130 |
|
131 |
-
prompt = re.sub("([^\x00-\x7F]|\w)(\.|\銆倈\?)", r"\1 \2
|
132 |
|
133 |
if normalize_text and language == "vi":
|
134 |
prompt = normalize_vietnamese_text(prompt)
|
135 |
|
136 |
-
print("
|
137 |
t0 = time.time()
|
138 |
out = MODEL.inference(
|
139 |
prompt,
|
@@ -145,100 +131,30 @@ def predict(
|
|
145 |
enable_text_splitting=True,
|
146 |
)
|
147 |
inference_time = time.time() - t0
|
148 |
-
|
149 |
-
metrics_text += (
|
150 |
-
f"Time to generate audio: {round(inference_time*1000)} milliseconds\n"
|
151 |
-
)
|
152 |
real_time_factor = (time.time() - t0) / out["wav"].shape[-1] * 24000
|
153 |
-
print(f"Real-time factor (RTF): {real_time_factor}")
|
154 |
metrics_text += f"Real-time factor (RTF): {real_time_factor:.2f}\n"
|
155 |
|
156 |
-
# Temporary hack for short sentences
|
157 |
keep_len = calculate_keep_len(prompt, language)
|
158 |
out["wav"] = out["wav"][:keep_len]
|
159 |
|
160 |
torchaudio.save("output.wav", torch.tensor(out["wav"]).unsqueeze(0), 24000)
|
161 |
|
162 |
except RuntimeError as e:
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
f"Exit due to: Unrecoverable exception caused by language:{language} prompt:{prompt}",
|
167 |
-
flush=True,
|
168 |
-
)
|
169 |
-
gr.Warning("Unhandled Exception encounter, please retry in a minute")
|
170 |
-
print("Cuda device-assert Runtime encountered need restart")
|
171 |
-
|
172 |
-
error_time = datetime.datetime.now().strftime("%d-%m-%Y-%H:%M:%S")
|
173 |
-
error_data = [
|
174 |
-
error_time,
|
175 |
-
prompt,
|
176 |
-
language,
|
177 |
-
audio_file_pth,
|
178 |
-
]
|
179 |
-
error_data = [str(e) if type(e) != str else e for e in error_data]
|
180 |
-
print(error_data)
|
181 |
-
print(speaker_wav)
|
182 |
-
write_io = StringIO()
|
183 |
-
csv.writer(write_io).writerows([error_data])
|
184 |
-
csv_upload = write_io.getvalue().encode()
|
185 |
-
|
186 |
-
filename = error_time + "_" + str(uuid.uuid4()) + ".csv"
|
187 |
-
print("Writing error csv")
|
188 |
-
error_api = HfApi()
|
189 |
-
error_api.upload_file(
|
190 |
-
path_or_fileobj=csv_upload,
|
191 |
-
path_in_repo=filename,
|
192 |
-
repo_id="coqui/xtts-flagged-dataset",
|
193 |
-
repo_type="dataset",
|
194 |
-
)
|
195 |
-
|
196 |
-
# speaker_wav
|
197 |
-
print("Writing error reference audio")
|
198 |
-
speaker_filename = error_time + "_reference_" + str(uuid.uuid4()) + ".wav"
|
199 |
-
error_api = HfApi()
|
200 |
-
error_api.upload_file(
|
201 |
-
path_or_fileobj=speaker_wav,
|
202 |
-
path_in_repo=speaker_filename,
|
203 |
-
repo_id="coqui/xtts-flagged-dataset",
|
204 |
-
repo_type="dataset",
|
205 |
-
)
|
206 |
|
207 |
-
|
208 |
-
space = api.get_space_runtime(repo_id=repo_id)
|
209 |
-
if space.stage != "BUILDING":
|
210 |
-
api.restart_space(repo_id=repo_id)
|
211 |
-
else:
|
212 |
-
print("TRIED TO RESTART but space is building")
|
213 |
-
|
214 |
-
else:
|
215 |
-
if "Failed to decode" in str(e):
|
216 |
-
print("Speaker encoding error", str(e))
|
217 |
-
metrics_text = gr.Warning(
|
218 |
-
metrics_text="It appears something wrong with reference, did you unmute your microphone?"
|
219 |
-
)
|
220 |
-
else:
|
221 |
-
print("RuntimeError: non device-side assert error:", str(e))
|
222 |
-
metrics_text = gr.Warning(
|
223 |
-
"Something unexpected happened please retry again."
|
224 |
-
)
|
225 |
-
return (None, metrics_text)
|
226 |
-
return ("output.wav", metrics_text)
|
227 |
|
228 |
|
229 |
title = "viXTTS Demo"
|
230 |
|
231 |
-
|
232 |
with gr.Blocks(analytics_enabled=False) as demo:
|
233 |
with gr.Row():
|
234 |
with gr.Column():
|
235 |
-
gr.Markdown(
|
236 |
-
"""
|
237 |
-
viXTTS Demo
|
238 |
-
"""
|
239 |
-
)
|
240 |
with gr.Column():
|
241 |
-
# placeholder to align the image
|
242 |
pass
|
243 |
|
244 |
with gr.Row():
|
@@ -251,33 +167,13 @@ with gr.Blocks(analytics_enabled=False) as demo:
|
|
251 |
language_gr = gr.Dropdown(
|
252 |
label="Language",
|
253 |
info="Select an output language for the synthesised speech",
|
254 |
-
choices=
|
255 |
-
"vi",
|
256 |
-
"en",
|
257 |
-
"es",
|
258 |
-
"fr",
|
259 |
-
"de",
|
260 |
-
"it",
|
261 |
-
"pt",
|
262 |
-
"pl",
|
263 |
-
"tr",
|
264 |
-
"ru",
|
265 |
-
"nl",
|
266 |
-
"cs",
|
267 |
-
"ar",
|
268 |
-
"zh-cn",
|
269 |
-
"ja",
|
270 |
-
"ko",
|
271 |
-
"hu",
|
272 |
-
"hi",
|
273 |
-
],
|
274 |
-
max_choices=1,
|
275 |
value="vi",
|
276 |
)
|
277 |
normalize_text = gr.Checkbox(
|
278 |
label="Normalize Vietnamese Text",
|
279 |
info="Normalize Vietnamese Text",
|
280 |
-
|
281 |
)
|
282 |
ref_gr = gr.Audio(
|
283 |
label="Reference Audio",
|
@@ -289,19 +185,14 @@ with gr.Blocks(analytics_enabled=False) as demo:
|
|
289 |
|
290 |
with gr.Column():
|
291 |
audio_gr = gr.Audio(label="Synthesised Audio", autoplay=True)
|
292 |
-
out_text_gr = gr.
|
293 |
|
294 |
tts_button.click(
|
295 |
predict,
|
296 |
-
[
|
297 |
-
input_text_gr,
|
298 |
-
language_gr,
|
299 |
-
ref_gr,
|
300 |
-
normalize_text,
|
301 |
-
],
|
302 |
outputs=[audio_gr, out_text_gr],
|
303 |
api_name="predict",
|
304 |
)
|
305 |
|
306 |
demo.queue()
|
307 |
-
demo.launch(debug=True, show_api=True
|
|
|
14 |
from TTS.tts.models.xtts import Xtts
|
15 |
from vinorm import TTSnorm
|
16 |
|
17 |
+
# Initialize Hugging Face API
|
|
|
|
|
18 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
19 |
api = HfApi(token=HF_TOKEN)
|
20 |
|
21 |
+
# Download model files if not already downloaded
|
22 |
+
print("Downloading viXTTS model files if not already present...")
|
23 |
checkpoint_dir = "model/"
|
24 |
repo_id = "capleaf/viXTTS"
|
25 |
use_deepspeed = False
|
|
|
40 |
local_dir=checkpoint_dir,
|
41 |
)
|
42 |
|
43 |
+
# Load model configuration and initialize model
|
44 |
xtts_config = os.path.join(checkpoint_dir, "config.json")
|
45 |
config = XttsConfig()
|
46 |
config.load_json(xtts_config)
|
|
|
51 |
if torch.cuda.is_available():
|
52 |
MODEL.cuda()
|
53 |
|
54 |
+
# Supported languages
|
55 |
supported_languages = config.languages
|
56 |
+
if "vi" not in supported_languages:
|
57 |
supported_languages.append("vi")
|
58 |
|
59 |
|
|
|
74 |
|
75 |
|
76 |
def calculate_keep_len(text, lang):
|
|
|
77 |
if lang in ["ja", "zh-cn"]:
|
78 |
return -1
|
79 |
|
|
|
87 |
return -1
|
88 |
|
89 |
|
90 |
+
def predict(prompt, language, audio_file_pth, normalize_text=True):
|
|
|
|
|
|
|
|
|
|
|
91 |
if language not in supported_languages:
|
92 |
metrics_text = gr.Warning(
|
93 |
+
f"Language {language} is not supported. Please choose from the dropdown."
|
94 |
)
|
95 |
+
return None, metrics_text
|
|
|
|
|
|
|
96 |
|
97 |
if len(prompt) < 2:
|
98 |
+
metrics_text = gr.Warning("Please provide a longer prompt text.")
|
99 |
+
return None, metrics_text
|
100 |
+
|
101 |
try:
|
102 |
metrics_text = ""
|
103 |
t_latent = time.time()
|
104 |
|
105 |
try:
|
106 |
+
gpt_cond_latent, speaker_embedding = MODEL.get_conditioning_latents(
|
107 |
+
audio_path=audio_file_pth,
|
|
|
|
|
|
|
108 |
gpt_cond_len=30,
|
109 |
gpt_cond_chunk_len=4,
|
110 |
max_ref_length=60,
|
111 |
)
|
|
|
112 |
except Exception as e:
|
113 |
+
print("Speaker encoding error:", str(e))
|
114 |
+
metrics_text = gr.Warning("Error with reference audio.")
|
115 |
+
return None, metrics_text
|
|
|
|
|
116 |
|
117 |
+
prompt = re.sub("([^\x00-\x7F]|\w)(\.|\銆倈\?)", r"\1 \2", prompt)
|
118 |
|
119 |
if normalize_text and language == "vi":
|
120 |
prompt = normalize_vietnamese_text(prompt)
|
121 |
|
122 |
+
print("Generating new audio...")
|
123 |
t0 = time.time()
|
124 |
out = MODEL.inference(
|
125 |
prompt,
|
|
|
131 |
enable_text_splitting=True,
|
132 |
)
|
133 |
inference_time = time.time() - t0
|
134 |
+
metrics_text += f"Time to generate audio: {round(inference_time * 1000)} ms\n"
|
|
|
|
|
|
|
135 |
real_time_factor = (time.time() - t0) / out["wav"].shape[-1] * 24000
|
|
|
136 |
metrics_text += f"Real-time factor (RTF): {real_time_factor:.2f}\n"
|
137 |
|
|
|
138 |
keep_len = calculate_keep_len(prompt, language)
|
139 |
out["wav"] = out["wav"][:keep_len]
|
140 |
|
141 |
torchaudio.save("output.wav", torch.tensor(out["wav"]).unsqueeze(0), 24000)
|
142 |
|
143 |
except RuntimeError as e:
|
144 |
+
print("RuntimeError:", str(e))
|
145 |
+
metrics_text = gr.Warning("An error occurred during processing.")
|
146 |
+
return None, metrics_text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
147 |
|
148 |
+
return "output.wav", metrics_text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
149 |
|
150 |
|
151 |
title = "viXTTS Demo"
|
152 |
|
|
|
153 |
with gr.Blocks(analytics_enabled=False) as demo:
|
154 |
with gr.Row():
|
155 |
with gr.Column():
|
156 |
+
gr.Markdown("## viXTTS Demo")
|
|
|
|
|
|
|
|
|
157 |
with gr.Column():
|
|
|
158 |
pass
|
159 |
|
160 |
with gr.Row():
|
|
|
167 |
language_gr = gr.Dropdown(
|
168 |
label="Language",
|
169 |
info="Select an output language for the synthesised speech",
|
170 |
+
choices=supported_languages,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
171 |
value="vi",
|
172 |
)
|
173 |
normalize_text = gr.Checkbox(
|
174 |
label="Normalize Vietnamese Text",
|
175 |
info="Normalize Vietnamese Text",
|
176 |
+
value=True,
|
177 |
)
|
178 |
ref_gr = gr.Audio(
|
179 |
label="Reference Audio",
|
|
|
185 |
|
186 |
with gr.Column():
|
187 |
audio_gr = gr.Audio(label="Synthesised Audio", autoplay=True)
|
188 |
+
out_text_gr = gr.Textbox(label="Metrics")
|
189 |
|
190 |
tts_button.click(
|
191 |
predict,
|
192 |
+
[input_text_gr, language_gr, ref_gr, normalize_text],
|
|
|
|
|
|
|
|
|
|
|
193 |
outputs=[audio_gr, out_text_gr],
|
194 |
api_name="predict",
|
195 |
)
|
196 |
|
197 |
demo.queue()
|
198 |
+
demo.launch(debug=True, show_api=True)
|