Spaces:
Running
Running
Sync from GitHub repo
Browse filesThis Space is synced from the GitHub repo: https://github.com/SWivid/F5-TTS. Please submit contributions to the Space there
- src/f5_tts/api.py +4 -0
- src/f5_tts/infer/utils_infer.py +17 -10
- src/f5_tts/train/finetune_cli.py +10 -3
- src/f5_tts/train/finetune_gradio.py +66 -44
src/f5_tts/api.py
CHANGED
@@ -15,6 +15,7 @@ from f5_tts.infer.utils_infer import (
|
|
15 |
preprocess_ref_audio_text,
|
16 |
remove_silence_for_generated_wav,
|
17 |
save_spectrogram,
|
|
|
18 |
target_sample_rate,
|
19 |
)
|
20 |
from f5_tts.model import DiT, UNetT
|
@@ -82,6 +83,9 @@ class F5TTS:
|
|
82 |
model_cls, model_cfg, ckpt_file, mel_spec_type, vocab_file, ode_method, use_ema, self.device
|
83 |
)
|
84 |
|
|
|
|
|
|
|
85 |
def export_wav(self, wav, file_wave, remove_silence=False):
|
86 |
sf.write(file_wave, wav, self.target_sample_rate)
|
87 |
|
|
|
15 |
preprocess_ref_audio_text,
|
16 |
remove_silence_for_generated_wav,
|
17 |
save_spectrogram,
|
18 |
+
transcribe,
|
19 |
target_sample_rate,
|
20 |
)
|
21 |
from f5_tts.model import DiT, UNetT
|
|
|
83 |
model_cls, model_cfg, ckpt_file, mel_spec_type, vocab_file, ode_method, use_ema, self.device
|
84 |
)
|
85 |
|
86 |
+
def transcribe(self, ref_audio, language=None):
|
87 |
+
return transcribe(ref_audio, language)
|
88 |
+
|
89 |
def export_wav(self, wav, file_wave, remove_silence=False):
|
90 |
sf.write(file_wave, wav, self.target_sample_rate)
|
91 |
|
src/f5_tts/infer/utils_infer.py
CHANGED
@@ -150,6 +150,22 @@ def initialize_asr_pipeline(device=device, dtype=None):
|
|
150 |
)
|
151 |
|
152 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
153 |
# load model checkpoint for inference
|
154 |
|
155 |
|
@@ -306,17 +322,8 @@ def preprocess_ref_audio_text(ref_audio_orig, ref_text, clip_short=True, show_in
|
|
306 |
show_info("Using cached reference text...")
|
307 |
ref_text = _ref_audio_cache[audio_hash]
|
308 |
else:
|
309 |
-
global asr_pipe
|
310 |
-
if asr_pipe is None:
|
311 |
-
initialize_asr_pipeline(device=device)
|
312 |
show_info("No reference text provided, transcribing reference audio...")
|
313 |
-
ref_text =
|
314 |
-
ref_audio,
|
315 |
-
chunk_length_s=30,
|
316 |
-
batch_size=128,
|
317 |
-
generate_kwargs={"task": "transcribe"},
|
318 |
-
return_timestamps=False,
|
319 |
-
)["text"].strip()
|
320 |
# Cache the transcribed text (not caching custom ref_text, enabling users to do manual tweak)
|
321 |
_ref_audio_cache[audio_hash] = ref_text
|
322 |
else:
|
|
|
150 |
)
|
151 |
|
152 |
|
153 |
+
# transcribe
|
154 |
+
|
155 |
+
|
156 |
+
def transcribe(ref_audio, language=None):
|
157 |
+
global asr_pipe
|
158 |
+
if asr_pipe is None:
|
159 |
+
initialize_asr_pipeline(device=device)
|
160 |
+
return asr_pipe(
|
161 |
+
ref_audio,
|
162 |
+
chunk_length_s=30,
|
163 |
+
batch_size=128,
|
164 |
+
generate_kwargs={"task": "transcribe", "language": language} if language else {"task": "transcribe"},
|
165 |
+
return_timestamps=False,
|
166 |
+
)["text"].strip()
|
167 |
+
|
168 |
+
|
169 |
# load model checkpoint for inference
|
170 |
|
171 |
|
|
|
322 |
show_info("Using cached reference text...")
|
323 |
ref_text = _ref_audio_cache[audio_hash]
|
324 |
else:
|
|
|
|
|
|
|
325 |
show_info("No reference text provided, transcribing reference audio...")
|
326 |
+
ref_text = transcribe(ref_audio)
|
|
|
|
|
|
|
|
|
|
|
|
|
327 |
# Cache the transcribed text (not caching custom ref_text, enabling users to do manual tweak)
|
328 |
_ref_audio_cache[audio_hash] = ref_text
|
329 |
else:
|
src/f5_tts/train/finetune_cli.py
CHANGED
@@ -13,6 +13,9 @@ from importlib.resources import files
|
|
13 |
target_sample_rate = 24000
|
14 |
n_mel_channels = 100
|
15 |
hop_length = 256
|
|
|
|
|
|
|
16 |
|
17 |
|
18 |
# -------------------------- Argument Parsing --------------------------- #
|
@@ -40,7 +43,7 @@ def parse_args():
|
|
40 |
parser.add_argument("--max_samples", type=int, default=64, help="Max sequences per batch")
|
41 |
parser.add_argument("--grad_accumulation_steps", type=int, default=1, help="Gradient accumulation steps")
|
42 |
parser.add_argument("--max_grad_norm", type=float, default=1.0, help="Max gradient norm for clipping")
|
43 |
-
parser.add_argument("--epochs", type=int, default=
|
44 |
parser.add_argument("--num_warmup_updates", type=int, default=300, help="Warmup steps")
|
45 |
parser.add_argument("--save_per_updates", type=int, default=10000, help="Save checkpoint every X steps")
|
46 |
parser.add_argument("--last_per_steps", type=int, default=50000, help="Save last checkpoint every X steps")
|
@@ -121,11 +124,15 @@ def main():
|
|
121 |
vocab_char_map, vocab_size = get_tokenizer(tokenizer_path, tokenizer)
|
122 |
|
123 |
print("\nvocab : ", vocab_size)
|
|
|
124 |
|
125 |
mel_spec_kwargs = dict(
|
126 |
-
|
127 |
-
n_mel_channels=n_mel_channels,
|
128 |
hop_length=hop_length,
|
|
|
|
|
|
|
|
|
129 |
)
|
130 |
|
131 |
model = CFM(
|
|
|
13 |
target_sample_rate = 24000
|
14 |
n_mel_channels = 100
|
15 |
hop_length = 256
|
16 |
+
win_length = 1024
|
17 |
+
n_fft = 1024
|
18 |
+
mel_spec_type = "vocos" # 'vocos' or 'bigvgan'
|
19 |
|
20 |
|
21 |
# -------------------------- Argument Parsing --------------------------- #
|
|
|
43 |
parser.add_argument("--max_samples", type=int, default=64, help="Max sequences per batch")
|
44 |
parser.add_argument("--grad_accumulation_steps", type=int, default=1, help="Gradient accumulation steps")
|
45 |
parser.add_argument("--max_grad_norm", type=float, default=1.0, help="Max gradient norm for clipping")
|
46 |
+
parser.add_argument("--epochs", type=int, default=100, help="Number of training epochs")
|
47 |
parser.add_argument("--num_warmup_updates", type=int, default=300, help="Warmup steps")
|
48 |
parser.add_argument("--save_per_updates", type=int, default=10000, help="Save checkpoint every X steps")
|
49 |
parser.add_argument("--last_per_steps", type=int, default=50000, help="Save last checkpoint every X steps")
|
|
|
124 |
vocab_char_map, vocab_size = get_tokenizer(tokenizer_path, tokenizer)
|
125 |
|
126 |
print("\nvocab : ", vocab_size)
|
127 |
+
print("\nvocoder : ", mel_spec_type)
|
128 |
|
129 |
mel_spec_kwargs = dict(
|
130 |
+
n_fft=n_fft,
|
|
|
131 |
hop_length=hop_length,
|
132 |
+
win_length=win_length,
|
133 |
+
n_mel_channels=n_mel_channels,
|
134 |
+
target_sample_rate=target_sample_rate,
|
135 |
+
mel_spec_type=mel_spec_type,
|
136 |
)
|
137 |
|
138 |
model = CFM(
|
src/f5_tts/train/finetune_gradio.py
CHANGED
@@ -26,12 +26,13 @@ from datasets import Dataset as Dataset_
|
|
26 |
from datasets.arrow_writer import ArrowWriter
|
27 |
from safetensors.torch import save_file
|
28 |
from scipy.io import wavfile
|
29 |
-
from transformers import pipeline
|
30 |
from cached_path import cached_path
|
31 |
from f5_tts.api import F5TTS
|
32 |
from f5_tts.model.utils import convert_char_to_pinyin
|
|
|
33 |
from importlib.resources import files
|
34 |
|
|
|
35 |
training_process = None
|
36 |
system = platform.system()
|
37 |
python_executable = sys.executable or "python"
|
@@ -47,8 +48,6 @@ file_train = "src/f5_tts/train/finetune_cli.py"
|
|
47 |
|
48 |
device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
|
49 |
|
50 |
-
pipe = None
|
51 |
-
|
52 |
|
53 |
# Save settings from a JSON file
|
54 |
def save_settings(
|
@@ -70,6 +69,7 @@ def save_settings(
|
|
70 |
tokenizer_file,
|
71 |
mixed_precision,
|
72 |
logger,
|
|
|
73 |
):
|
74 |
path_project = os.path.join(path_project_ckpts, project_name)
|
75 |
os.makedirs(path_project, exist_ok=True)
|
@@ -93,6 +93,7 @@ def save_settings(
|
|
93 |
"tokenizer_file": tokenizer_file,
|
94 |
"mixed_precision": mixed_precision,
|
95 |
"logger": logger,
|
|
|
96 |
}
|
97 |
with open(file_setting, "w") as f:
|
98 |
json.dump(settings, f, indent=4)
|
@@ -124,6 +125,7 @@ def load_settings(project_name):
|
|
124 |
"tokenizer_file": "",
|
125 |
"mixed_precision": "none",
|
126 |
"logger": "wandb",
|
|
|
127 |
}
|
128 |
return (
|
129 |
settings["exp_name"],
|
@@ -143,12 +145,15 @@ def load_settings(project_name):
|
|
143 |
settings["tokenizer_file"],
|
144 |
settings["mixed_precision"],
|
145 |
settings["logger"],
|
|
|
146 |
)
|
147 |
|
148 |
with open(file_setting, "r") as f:
|
149 |
settings = json.load(f)
|
150 |
if "logger" not in settings:
|
151 |
settings["logger"] = "wandb"
|
|
|
|
|
152 |
return (
|
153 |
settings["exp_name"],
|
154 |
settings["learning_rate"],
|
@@ -167,6 +172,7 @@ def load_settings(project_name):
|
|
167 |
settings["tokenizer_file"],
|
168 |
settings["mixed_precision"],
|
169 |
settings["logger"],
|
|
|
170 |
)
|
171 |
|
172 |
|
@@ -381,18 +387,17 @@ def start_training(
|
|
381 |
mixed_precision="fp16",
|
382 |
stream=False,
|
383 |
logger="wandb",
|
|
|
384 |
):
|
385 |
-
global training_process, tts_api, stop_signal
|
386 |
|
387 |
-
if tts_api is not None
|
388 |
if tts_api is not None:
|
389 |
del tts_api
|
390 |
-
|
391 |
-
del pipe
|
392 |
gc.collect()
|
393 |
torch.cuda.empty_cache()
|
394 |
tts_api = None
|
395 |
-
pipe = None
|
396 |
|
397 |
path_project = os.path.join(path_data, dataset_name)
|
398 |
|
@@ -447,11 +452,10 @@ def start_training(
|
|
447 |
f"--dataset_name {dataset_name}"
|
448 |
)
|
449 |
|
450 |
-
|
451 |
-
cmd += f" --finetune {finetune}"
|
452 |
|
453 |
if file_checkpoint_train != "":
|
454 |
-
cmd += f" --
|
455 |
|
456 |
if tokenizer_file != "":
|
457 |
cmd += f" --tokenizer_path {tokenizer_file}"
|
@@ -460,7 +464,10 @@ def start_training(
|
|
460 |
|
461 |
cmd += f" --log_samples True --logger {logger} "
|
462 |
|
463 |
-
|
|
|
|
|
|
|
464 |
|
465 |
save_settings(
|
466 |
dataset_name,
|
@@ -481,6 +488,7 @@ def start_training(
|
|
481 |
tokenizer_file,
|
482 |
mixed_precision,
|
483 |
logger,
|
|
|
484 |
)
|
485 |
|
486 |
try:
|
@@ -641,27 +649,6 @@ def create_data_project(name, tokenizer_type):
|
|
641 |
return gr.update(choices=project_list, value=name)
|
642 |
|
643 |
|
644 |
-
def transcribe(file_audio, language="english"):
|
645 |
-
global pipe
|
646 |
-
|
647 |
-
if pipe is None:
|
648 |
-
pipe = pipeline(
|
649 |
-
"automatic-speech-recognition",
|
650 |
-
model="openai/whisper-large-v3-turbo",
|
651 |
-
torch_dtype=torch.float16,
|
652 |
-
device=device,
|
653 |
-
)
|
654 |
-
|
655 |
-
text_transcribe = pipe(
|
656 |
-
file_audio,
|
657 |
-
chunk_length_s=30,
|
658 |
-
batch_size=128,
|
659 |
-
generate_kwargs={"task": "transcribe", "language": language},
|
660 |
-
return_timestamps=False,
|
661 |
-
)["text"].strip()
|
662 |
-
return text_transcribe
|
663 |
-
|
664 |
-
|
665 |
def transcribe_all(name_project, audio_files, language, user=False, progress=gr.Progress()):
|
666 |
path_project = os.path.join(path_data, name_project)
|
667 |
path_dataset = os.path.join(path_project, "dataset")
|
@@ -758,11 +745,9 @@ def get_correct_audio_path(
|
|
758 |
# Case 2: If it has a supported extension but is not a full path
|
759 |
elif has_supported_extension(audio_input) and not os.path.isabs(audio_input):
|
760 |
file_audio = os.path.join(base_path, audio_input)
|
761 |
-
print("2")
|
762 |
|
763 |
# Case 3: If only the name is given (no extension and not a full path)
|
764 |
elif not has_supported_extension(audio_input) and not os.path.isabs(audio_input):
|
765 |
-
print("3")
|
766 |
for ext in supported_formats:
|
767 |
potential_file = os.path.join(base_path, f"{audio_input}.{ext}")
|
768 |
if os.path.exists(potential_file):
|
@@ -816,9 +801,12 @@ def create_metadata(name_project, ch_tokenizer, progress=gr.Progress()):
|
|
816 |
continue
|
817 |
|
818 |
if duration < 1 or duration > 25:
|
819 |
-
|
|
|
|
|
|
|
820 |
continue
|
821 |
-
if len(text) <
|
822 |
error_files.append([file_audio, "very small text len 3"])
|
823 |
continue
|
824 |
|
@@ -1208,7 +1196,9 @@ def get_random_sample_infer(project_name):
|
|
1208 |
)
|
1209 |
|
1210 |
|
1211 |
-
def infer(
|
|
|
|
|
1212 |
global last_checkpoint, last_device, tts_api, last_ema
|
1213 |
|
1214 |
if not os.path.isfile(file_checkpoint):
|
@@ -1238,8 +1228,17 @@ def infer(project, file_checkpoint, exp_name, ref_text, ref_audio, gen_text, nfe
|
|
1238 |
print("update >> ", device_test, file_checkpoint, use_ema)
|
1239 |
|
1240 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f:
|
1241 |
-
tts_api.infer(
|
1242 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1243 |
|
1244 |
|
1245 |
def check_finetune(finetune):
|
@@ -1506,6 +1505,7 @@ Skip this step if you have your dataset, raw.arrow, duration.json, and vocab.txt
|
|
1506 |
```"""
|
1507 |
)
|
1508 |
ch_tokenizern = gr.Checkbox(label="Create Vocabulary", value=False, visible=False)
|
|
|
1509 |
bt_prepare = bt_create = gr.Button("Prepare")
|
1510 |
txt_info_prepare = gr.Text(label="Info", value="")
|
1511 |
txt_vocab_prepare = gr.Text(label="Vocab", value="")
|
@@ -1560,6 +1560,7 @@ If you encounter a memory error, try reducing the batch size per GPU to a smalle
|
|
1560 |
last_per_steps = gr.Number(label="Last per Steps", value=100)
|
1561 |
|
1562 |
with gr.Row():
|
|
|
1563 |
mixed_precision = gr.Radio(label="mixed_precision", choices=["none", "fp16", "bf16"], value="none")
|
1564 |
cd_logger = gr.Radio(label="logger", choices=["wandb", "tensorboard"], value="wandb")
|
1565 |
start_button = gr.Button("Start Training")
|
@@ -1584,6 +1585,7 @@ If you encounter a memory error, try reducing the batch size per GPU to a smalle
|
|
1584 |
tokenizer_filev,
|
1585 |
mixed_precisionv,
|
1586 |
cd_loggerv,
|
|
|
1587 |
) = load_settings(projects_selelect)
|
1588 |
exp_name.value = exp_namev
|
1589 |
learning_rate.value = learning_ratev
|
@@ -1602,6 +1604,7 @@ If you encounter a memory error, try reducing the batch size per GPU to a smalle
|
|
1602 |
tokenizer_file.value = tokenizer_filev
|
1603 |
mixed_precision.value = mixed_precisionv
|
1604 |
cd_logger.value = cd_loggerv
|
|
|
1605 |
|
1606 |
ch_stream = gr.Checkbox(label="Stream Output Experiment", value=True)
|
1607 |
txt_info_train = gr.Text(label="Info", value="")
|
@@ -1660,6 +1663,7 @@ If you encounter a memory error, try reducing the batch size per GPU to a smalle
|
|
1660 |
mixed_precision,
|
1661 |
ch_stream,
|
1662 |
cd_logger,
|
|
|
1663 |
],
|
1664 |
outputs=[txt_info_train, start_button, stop_button],
|
1665 |
)
|
@@ -1732,12 +1736,17 @@ If you encounter a memory error, try reducing the batch size per GPU to a smalle
|
|
1732 |
|
1733 |
with gr.TabItem("Test Model"):
|
1734 |
gr.Markdown("""```plaintext
|
1735 |
-
SOS: Check the use_ema setting (True or False) for your model to see what works best for you.
|
1736 |
```""")
|
1737 |
exp_name = gr.Radio(label="Model", choices=["F5-TTS", "E2-TTS"], value="F5-TTS")
|
1738 |
list_checkpoints, checkpoint_select = get_checkpoints_project(projects_selelect, False)
|
1739 |
|
1740 |
-
|
|
|
|
|
|
|
|
|
|
|
1741 |
ch_use_ema = gr.Checkbox(label="Use EMA", value=True)
|
1742 |
with gr.Row():
|
1743 |
cm_checkpoint = gr.Dropdown(
|
@@ -1757,14 +1766,27 @@ SOS: Check the use_ema setting (True or False) for your model to see what works
|
|
1757 |
|
1758 |
with gr.Row():
|
1759 |
txt_info_gpu = gr.Textbox("", label="Device")
|
|
|
1760 |
check_button_infer = gr.Button("Infer")
|
1761 |
|
1762 |
gen_audio = gr.Audio(label="Audio Gen", type="filepath")
|
1763 |
|
1764 |
check_button_infer.click(
|
1765 |
fn=infer,
|
1766 |
-
inputs=[
|
1767 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1768 |
)
|
1769 |
|
1770 |
bt_checkpoint_refresh.click(fn=get_checkpoints_project, inputs=[cm_project], outputs=[cm_checkpoint])
|
|
|
26 |
from datasets.arrow_writer import ArrowWriter
|
27 |
from safetensors.torch import save_file
|
28 |
from scipy.io import wavfile
|
|
|
29 |
from cached_path import cached_path
|
30 |
from f5_tts.api import F5TTS
|
31 |
from f5_tts.model.utils import convert_char_to_pinyin
|
32 |
+
from f5_tts.infer.utils_infer import transcribe
|
33 |
from importlib.resources import files
|
34 |
|
35 |
+
|
36 |
training_process = None
|
37 |
system = platform.system()
|
38 |
python_executable = sys.executable or "python"
|
|
|
48 |
|
49 |
device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
|
50 |
|
|
|
|
|
51 |
|
52 |
# Save settings from a JSON file
|
53 |
def save_settings(
|
|
|
69 |
tokenizer_file,
|
70 |
mixed_precision,
|
71 |
logger,
|
72 |
+
ch_8bit_adam,
|
73 |
):
|
74 |
path_project = os.path.join(path_project_ckpts, project_name)
|
75 |
os.makedirs(path_project, exist_ok=True)
|
|
|
93 |
"tokenizer_file": tokenizer_file,
|
94 |
"mixed_precision": mixed_precision,
|
95 |
"logger": logger,
|
96 |
+
"bnb_optimizer": ch_8bit_adam,
|
97 |
}
|
98 |
with open(file_setting, "w") as f:
|
99 |
json.dump(settings, f, indent=4)
|
|
|
125 |
"tokenizer_file": "",
|
126 |
"mixed_precision": "none",
|
127 |
"logger": "wandb",
|
128 |
+
"bnb_optimizer": False,
|
129 |
}
|
130 |
return (
|
131 |
settings["exp_name"],
|
|
|
145 |
settings["tokenizer_file"],
|
146 |
settings["mixed_precision"],
|
147 |
settings["logger"],
|
148 |
+
settings["bnb_optimizer"],
|
149 |
)
|
150 |
|
151 |
with open(file_setting, "r") as f:
|
152 |
settings = json.load(f)
|
153 |
if "logger" not in settings:
|
154 |
settings["logger"] = "wandb"
|
155 |
+
if "bnb_optimizer" not in settings:
|
156 |
+
settings["bnb_optimizer"] = False
|
157 |
return (
|
158 |
settings["exp_name"],
|
159 |
settings["learning_rate"],
|
|
|
172 |
settings["tokenizer_file"],
|
173 |
settings["mixed_precision"],
|
174 |
settings["logger"],
|
175 |
+
settings["bnb_optimizer"],
|
176 |
)
|
177 |
|
178 |
|
|
|
387 |
mixed_precision="fp16",
|
388 |
stream=False,
|
389 |
logger="wandb",
|
390 |
+
ch_8bit_adam=False,
|
391 |
):
|
392 |
+
global training_process, tts_api, stop_signal
|
393 |
|
394 |
+
if tts_api is not None:
|
395 |
if tts_api is not None:
|
396 |
del tts_api
|
397 |
+
|
|
|
398 |
gc.collect()
|
399 |
torch.cuda.empty_cache()
|
400 |
tts_api = None
|
|
|
401 |
|
402 |
path_project = os.path.join(path_data, dataset_name)
|
403 |
|
|
|
452 |
f"--dataset_name {dataset_name}"
|
453 |
)
|
454 |
|
455 |
+
cmd += f" --finetune {finetune}"
|
|
|
456 |
|
457 |
if file_checkpoint_train != "":
|
458 |
+
cmd += f" --pretrain {file_checkpoint_train}"
|
459 |
|
460 |
if tokenizer_file != "":
|
461 |
cmd += f" --tokenizer_path {tokenizer_file}"
|
|
|
464 |
|
465 |
cmd += f" --log_samples True --logger {logger} "
|
466 |
|
467 |
+
if ch_8bit_adam:
|
468 |
+
cmd += " --bnb_optimizer True "
|
469 |
+
|
470 |
+
print("run command : \n" + cmd + "\n")
|
471 |
|
472 |
save_settings(
|
473 |
dataset_name,
|
|
|
488 |
tokenizer_file,
|
489 |
mixed_precision,
|
490 |
logger,
|
491 |
+
ch_8bit_adam,
|
492 |
)
|
493 |
|
494 |
try:
|
|
|
649 |
return gr.update(choices=project_list, value=name)
|
650 |
|
651 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
652 |
def transcribe_all(name_project, audio_files, language, user=False, progress=gr.Progress()):
|
653 |
path_project = os.path.join(path_data, name_project)
|
654 |
path_dataset = os.path.join(path_project, "dataset")
|
|
|
745 |
# Case 2: If it has a supported extension but is not a full path
|
746 |
elif has_supported_extension(audio_input) and not os.path.isabs(audio_input):
|
747 |
file_audio = os.path.join(base_path, audio_input)
|
|
|
748 |
|
749 |
# Case 3: If only the name is given (no extension and not a full path)
|
750 |
elif not has_supported_extension(audio_input) and not os.path.isabs(audio_input):
|
|
|
751 |
for ext in supported_formats:
|
752 |
potential_file = os.path.join(base_path, f"{audio_input}.{ext}")
|
753 |
if os.path.exists(potential_file):
|
|
|
801 |
continue
|
802 |
|
803 |
if duration < 1 or duration > 25:
|
804 |
+
if duration > 25:
|
805 |
+
error_files.append([file_audio, "duration > 25 sec"])
|
806 |
+
if duration < 1:
|
807 |
+
error_files.append([file_audio, "duration < 1 sec "])
|
808 |
continue
|
809 |
+
if len(text) < 3:
|
810 |
error_files.append([file_audio, "very small text len 3"])
|
811 |
continue
|
812 |
|
|
|
1196 |
)
|
1197 |
|
1198 |
|
1199 |
+
def infer(
|
1200 |
+
project, file_checkpoint, exp_name, ref_text, ref_audio, gen_text, nfe_step, use_ema, speed, seed, remove_silence
|
1201 |
+
):
|
1202 |
global last_checkpoint, last_device, tts_api, last_ema
|
1203 |
|
1204 |
if not os.path.isfile(file_checkpoint):
|
|
|
1228 |
print("update >> ", device_test, file_checkpoint, use_ema)
|
1229 |
|
1230 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f:
|
1231 |
+
tts_api.infer(
|
1232 |
+
gen_text=gen_text.lower().strip(),
|
1233 |
+
ref_text=ref_text.lower().strip(),
|
1234 |
+
ref_file=ref_audio,
|
1235 |
+
nfe_step=nfe_step,
|
1236 |
+
file_wave=f.name,
|
1237 |
+
speed=speed,
|
1238 |
+
seed=seed,
|
1239 |
+
remove_silence=remove_silence,
|
1240 |
+
)
|
1241 |
+
return f.name, tts_api.device, str(tts_api.seed)
|
1242 |
|
1243 |
|
1244 |
def check_finetune(finetune):
|
|
|
1505 |
```"""
|
1506 |
)
|
1507 |
ch_tokenizern = gr.Checkbox(label="Create Vocabulary", value=False, visible=False)
|
1508 |
+
|
1509 |
bt_prepare = bt_create = gr.Button("Prepare")
|
1510 |
txt_info_prepare = gr.Text(label="Info", value="")
|
1511 |
txt_vocab_prepare = gr.Text(label="Vocab", value="")
|
|
|
1560 |
last_per_steps = gr.Number(label="Last per Steps", value=100)
|
1561 |
|
1562 |
with gr.Row():
|
1563 |
+
ch_8bit_adam = gr.Checkbox(label="Use 8-bit Adam optimizer")
|
1564 |
mixed_precision = gr.Radio(label="mixed_precision", choices=["none", "fp16", "bf16"], value="none")
|
1565 |
cd_logger = gr.Radio(label="logger", choices=["wandb", "tensorboard"], value="wandb")
|
1566 |
start_button = gr.Button("Start Training")
|
|
|
1585 |
tokenizer_filev,
|
1586 |
mixed_precisionv,
|
1587 |
cd_loggerv,
|
1588 |
+
ch_8bit_adamv,
|
1589 |
) = load_settings(projects_selelect)
|
1590 |
exp_name.value = exp_namev
|
1591 |
learning_rate.value = learning_ratev
|
|
|
1604 |
tokenizer_file.value = tokenizer_filev
|
1605 |
mixed_precision.value = mixed_precisionv
|
1606 |
cd_logger.value = cd_loggerv
|
1607 |
+
ch_8bit_adam.value = ch_8bit_adamv
|
1608 |
|
1609 |
ch_stream = gr.Checkbox(label="Stream Output Experiment", value=True)
|
1610 |
txt_info_train = gr.Text(label="Info", value="")
|
|
|
1663 |
mixed_precision,
|
1664 |
ch_stream,
|
1665 |
cd_logger,
|
1666 |
+
ch_8bit_adam,
|
1667 |
],
|
1668 |
outputs=[txt_info_train, start_button, stop_button],
|
1669 |
)
|
|
|
1736 |
|
1737 |
with gr.TabItem("Test Model"):
|
1738 |
gr.Markdown("""```plaintext
|
1739 |
+
SOS: Check the use_ema setting (True or False) for your model to see what works best for you. use seed -1 from random
|
1740 |
```""")
|
1741 |
exp_name = gr.Radio(label="Model", choices=["F5-TTS", "E2-TTS"], value="F5-TTS")
|
1742 |
list_checkpoints, checkpoint_select = get_checkpoints_project(projects_selelect, False)
|
1743 |
|
1744 |
+
with gr.Row():
|
1745 |
+
nfe_step = gr.Number(label="NFE Step", value=32)
|
1746 |
+
speed = gr.Slider(label="Speed", value=1.0, minimum=0.3, maximum=2.0, step=0.1)
|
1747 |
+
seed = gr.Number(label="Seed", value=-1, minimum=-1)
|
1748 |
+
remove_silence = gr.Checkbox(label="Remove Silence")
|
1749 |
+
|
1750 |
ch_use_ema = gr.Checkbox(label="Use EMA", value=True)
|
1751 |
with gr.Row():
|
1752 |
cm_checkpoint = gr.Dropdown(
|
|
|
1766 |
|
1767 |
with gr.Row():
|
1768 |
txt_info_gpu = gr.Textbox("", label="Device")
|
1769 |
+
seed_info = gr.Text(label="Seed :")
|
1770 |
check_button_infer = gr.Button("Infer")
|
1771 |
|
1772 |
gen_audio = gr.Audio(label="Audio Gen", type="filepath")
|
1773 |
|
1774 |
check_button_infer.click(
|
1775 |
fn=infer,
|
1776 |
+
inputs=[
|
1777 |
+
cm_project,
|
1778 |
+
cm_checkpoint,
|
1779 |
+
exp_name,
|
1780 |
+
ref_text,
|
1781 |
+
ref_audio,
|
1782 |
+
gen_text,
|
1783 |
+
nfe_step,
|
1784 |
+
ch_use_ema,
|
1785 |
+
speed,
|
1786 |
+
seed,
|
1787 |
+
remove_silence,
|
1788 |
+
],
|
1789 |
+
outputs=[gen_audio, txt_info_gpu, seed_info],
|
1790 |
)
|
1791 |
|
1792 |
bt_checkpoint_refresh.click(fn=get_checkpoints_project, inputs=[cm_project], outputs=[cm_checkpoint])
|