Spaces:
Running
on
Zero
Running
on
Zero
Commit
•
4ca6628
1
Parent(s):
0f3d0d9
Update app.py (#18)
Browse files- Update app.py (520692dd61a5947d0838a926d2856581f3b3c5bc)
Co-authored-by: Yushen CHEN <[email protected]>
app.py
CHANGED
@@ -8,7 +8,7 @@ import tempfile
|
|
8 |
from einops import rearrange
|
9 |
from ema_pytorch import EMA
|
10 |
from vocos import Vocos
|
11 |
-
from pydub import AudioSegment
|
12 |
from model import CFM, UNetT, DiT, MMDiT
|
13 |
from cached_path import cached_path
|
14 |
from model.utils import (
|
@@ -19,6 +19,7 @@ from model.utils import (
|
|
19 |
from transformers import pipeline
|
20 |
import spaces
|
21 |
import librosa
|
|
|
22 |
from txtsplit import txtsplit
|
23 |
from detoxify import Detoxify
|
24 |
|
@@ -49,8 +50,8 @@ speed = 1.0
|
|
49 |
# fix_duration = 27 # None or float (duration in seconds)
|
50 |
fix_duration = None
|
51 |
|
52 |
-
def load_model(exp_name, model_cls, model_cfg, ckpt_step):
|
53 |
-
checkpoint = torch.load(str(cached_path(f"hf://SWivid/
|
54 |
vocab_char_map, vocab_size = get_tokenizer("Emilia_ZH_EN", "pinyin")
|
55 |
model = CFM(
|
56 |
transformer=model_cls(
|
@@ -73,14 +74,14 @@ def load_model(exp_name, model_cls, model_cfg, ckpt_step):
|
|
73 |
ema_model.load_state_dict(checkpoint['ema_model_state_dict'])
|
74 |
ema_model.copy_params_from_ema_to_model()
|
75 |
|
76 |
-
return
|
77 |
|
78 |
# load models
|
79 |
F5TTS_model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
|
80 |
E2TTS_model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4)
|
81 |
|
82 |
-
F5TTS_ema_model
|
83 |
-
E2TTS_ema_model
|
84 |
|
85 |
@spaces.GPU
|
86 |
def infer(ref_audio_orig, ref_text, gen_text, exp_name, remove_silence, progress = gr.Progress()):
|
@@ -91,6 +92,12 @@ def infer(ref_audio_orig, ref_text, gen_text, exp_name, remove_silence, progress
|
|
91 |
gr.Info("Converting audio...")
|
92 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f:
|
93 |
aseg = AudioSegment.from_file(ref_audio_orig)
|
|
|
|
|
|
|
|
|
|
|
|
|
94 |
# Convert to mono
|
95 |
aseg = aseg.set_channels(1)
|
96 |
audio_duration = len(aseg)
|
@@ -101,10 +108,8 @@ def infer(ref_audio_orig, ref_text, gen_text, exp_name, remove_silence, progress
|
|
101 |
ref_audio = f.name
|
102 |
if exp_name == "F5-TTS":
|
103 |
ema_model = F5TTS_ema_model
|
104 |
-
base_model = F5TTS_base_model
|
105 |
elif exp_name == "E2-TTS":
|
106 |
ema_model = E2TTS_ema_model
|
107 |
-
base_model = E2TTS_base_model
|
108 |
|
109 |
if not ref_text.strip():
|
110 |
gr.Info("No reference text provided, transcribing reference audio...")
|
@@ -119,6 +124,7 @@ def infer(ref_audio_orig, ref_text, gen_text, exp_name, remove_silence, progress
|
|
119 |
else:
|
120 |
gr.Info("Using custom reference text...")
|
121 |
audio, sr = torchaudio.load(ref_audio)
|
|
|
122 |
# Audio
|
123 |
if audio.shape[0] > 1:
|
124 |
audio = torch.mean(audio, dim=0, keepdim=True)
|
@@ -130,7 +136,7 @@ def infer(ref_audio_orig, ref_text, gen_text, exp_name, remove_silence, progress
|
|
130 |
audio = resampler(audio)
|
131 |
audio = audio.to(device)
|
132 |
# Chunk
|
133 |
-
chunks = txtsplit(gen_text,
|
134 |
results = []
|
135 |
generated_mel_specs = []
|
136 |
for chunk in progress.tqdm(chunks):
|
@@ -144,14 +150,14 @@ def infer(ref_audio_orig, ref_text, gen_text, exp_name, remove_silence, progress
|
|
144 |
# duration = int(fix_duration * target_sample_rate / hop_length)
|
145 |
# else:
|
146 |
zh_pause_punc = r"。,、;:?!"
|
147 |
-
ref_text_len = len(ref_text) + len(re.findall(zh_pause_punc, ref_text))
|
148 |
-
|
149 |
duration = ref_audio_len + int(ref_audio_len / ref_text_len * gen_text_len / speed)
|
150 |
|
151 |
# inference
|
152 |
gr.Info(f"Generating audio using {exp_name}")
|
153 |
with torch.inference_mode():
|
154 |
-
generated, _ =
|
155 |
cond=audio,
|
156 |
text=final_text_list,
|
157 |
duration=duration,
|
@@ -174,12 +180,23 @@ def infer(ref_audio_orig, ref_text, gen_text, exp_name, remove_silence, progress
|
|
174 |
generated_wave = np.concatenate(results)
|
175 |
if remove_silence:
|
176 |
gr.Info("Removing audio silences... This may take a moment")
|
177 |
-
non_silent_intervals = librosa.effects.split(generated_wave, top_db=30)
|
178 |
-
non_silent_wave = np.array([])
|
179 |
-
for interval in non_silent_intervals:
|
180 |
-
|
181 |
-
|
182 |
-
generated_wave = non_silent_wave
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
183 |
|
184 |
|
185 |
# spectogram
|
|
|
8 |
from einops import rearrange
|
9 |
from ema_pytorch import EMA
|
10 |
from vocos import Vocos
|
11 |
+
from pydub import AudioSegment, silence
|
12 |
from model import CFM, UNetT, DiT, MMDiT
|
13 |
from cached_path import cached_path
|
14 |
from model.utils import (
|
|
|
19 |
from transformers import pipeline
|
20 |
import spaces
|
21 |
import librosa
|
22 |
+
import soundfile as sf
|
23 |
from txtsplit import txtsplit
|
24 |
from detoxify import Detoxify
|
25 |
|
|
|
50 |
# fix_duration = 27 # None or float (duration in seconds)
|
51 |
fix_duration = None
|
52 |
|
53 |
+
def load_model(repo_name, exp_name, model_cls, model_cfg, ckpt_step):
|
54 |
+
checkpoint = torch.load(str(cached_path(f"hf://SWivid/{repo_name}/{exp_name}/model_{ckpt_step}.pt")), map_location=device)
|
55 |
vocab_char_map, vocab_size = get_tokenizer("Emilia_ZH_EN", "pinyin")
|
56 |
model = CFM(
|
57 |
transformer=model_cls(
|
|
|
74 |
ema_model.load_state_dict(checkpoint['ema_model_state_dict'])
|
75 |
ema_model.copy_params_from_ema_to_model()
|
76 |
|
77 |
+
return model
|
78 |
|
79 |
# load models
|
80 |
F5TTS_model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
|
81 |
E2TTS_model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4)
|
82 |
|
83 |
+
F5TTS_ema_model = load_model("F5-TTS", "F5TTS_Base", DiT, F5TTS_model_cfg, 1200000)
|
84 |
+
E2TTS_ema_model = load_model("E2-TTS", "E2TTS_Base", UNetT, E2TTS_model_cfg, 1200000)
|
85 |
|
86 |
@spaces.GPU
|
87 |
def infer(ref_audio_orig, ref_text, gen_text, exp_name, remove_silence, progress = gr.Progress()):
|
|
|
92 |
gr.Info("Converting audio...")
|
93 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f:
|
94 |
aseg = AudioSegment.from_file(ref_audio_orig)
|
95 |
+
# remove long silence in reference audio
|
96 |
+
non_silent_segs = silence.split_on_silence(aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=500)
|
97 |
+
non_silent_wave = AudioSegment.silent(duration=0)
|
98 |
+
for non_silent_seg in non_silent_segs:
|
99 |
+
non_silent_wave += non_silent_seg
|
100 |
+
aseg = non_silent_wave
|
101 |
# Convert to mono
|
102 |
aseg = aseg.set_channels(1)
|
103 |
audio_duration = len(aseg)
|
|
|
108 |
ref_audio = f.name
|
109 |
if exp_name == "F5-TTS":
|
110 |
ema_model = F5TTS_ema_model
|
|
|
111 |
elif exp_name == "E2-TTS":
|
112 |
ema_model = E2TTS_ema_model
|
|
|
113 |
|
114 |
if not ref_text.strip():
|
115 |
gr.Info("No reference text provided, transcribing reference audio...")
|
|
|
124 |
else:
|
125 |
gr.Info("Using custom reference text...")
|
126 |
audio, sr = torchaudio.load(ref_audio)
|
127 |
+
max_chars = int(len(ref_text) / (audio.shape[-1] / sr) * (30 - audio.shape[-1] / sr))
|
128 |
# Audio
|
129 |
if audio.shape[0] > 1:
|
130 |
audio = torch.mean(audio, dim=0, keepdim=True)
|
|
|
136 |
audio = resampler(audio)
|
137 |
audio = audio.to(device)
|
138 |
# Chunk
|
139 |
+
chunks = txtsplit(gen_text, 0.7*max_chars, 0.9*max_chars)
|
140 |
results = []
|
141 |
generated_mel_specs = []
|
142 |
for chunk in progress.tqdm(chunks):
|
|
|
150 |
# duration = int(fix_duration * target_sample_rate / hop_length)
|
151 |
# else:
|
152 |
zh_pause_punc = r"。,、;:?!"
|
153 |
+
ref_text_len = len(ref_text.encode('utf-8')) + 3 * len(re.findall(zh_pause_punc, ref_text))
|
154 |
+
chunk = len(chunk.encode('utf-8')) + 3 * len(re.findall(zh_pause_punc, gen_text))
|
155 |
duration = ref_audio_len + int(ref_audio_len / ref_text_len * gen_text_len / speed)
|
156 |
|
157 |
# inference
|
158 |
gr.Info(f"Generating audio using {exp_name}")
|
159 |
with torch.inference_mode():
|
160 |
+
generated, _ = ema_model.sample(
|
161 |
cond=audio,
|
162 |
text=final_text_list,
|
163 |
duration=duration,
|
|
|
180 |
generated_wave = np.concatenate(results)
|
181 |
if remove_silence:
|
182 |
gr.Info("Removing audio silences... This may take a moment")
|
183 |
+
# non_silent_intervals = librosa.effects.split(generated_wave, top_db=30)
|
184 |
+
# non_silent_wave = np.array([])
|
185 |
+
# for interval in non_silent_intervals:
|
186 |
+
# start, end = interval
|
187 |
+
# non_silent_wave = np.concatenate([non_silent_wave, generated_wave[start:end]])
|
188 |
+
# generated_wave = non_silent_wave
|
189 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f:
|
190 |
+
sf.write(f.name, generated_wave, target_sample_rate)
|
191 |
+
aseg = AudioSegment.from_file(f.name)
|
192 |
+
non_silent_segs = silence.split_on_silence(aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=500)
|
193 |
+
non_silent_wave = AudioSegment.silent(duration=0)
|
194 |
+
for non_silent_seg in non_silent_segs:
|
195 |
+
non_silent_wave += non_silent_seg
|
196 |
+
aseg = non_silent_wave
|
197 |
+
aseg.export(f.name, format="wav")
|
198 |
+
generated_wave, _ = torchaudio.load(f.name)
|
199 |
+
generated_wave = generated_wave.squeeze().cpu().numpy()
|
200 |
|
201 |
|
202 |
# spectogram
|