Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,117 +1,350 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
import os
|
4 |
import torch
|
|
|
|
|
|
|
|
|
5 |
from pathlib import Path
|
6 |
from huggingface_hub import hf_hub_download
|
7 |
-
import
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
|
9 |
-
def
|
10 |
-
"""
|
11 |
-
|
12 |
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
# Check if it's a HuggingFace URL
|
20 |
-
if checkpoint_path.startswith("hf://"):
|
21 |
-
# Parse HF URL: hf://username/repo/path/to/model.pt
|
22 |
-
match = re.match(r"hf://([^/]+/[^/]+)/(.+)", checkpoint_path)
|
23 |
-
if match:
|
24 |
-
repo_id = match.group(1)
|
25 |
-
filename = match.group(2)
|
26 |
-
|
27 |
-
print(f"Loading from HuggingFace: {repo_id}/{filename}")
|
28 |
-
|
29 |
-
# Download from HuggingFace
|
30 |
-
local_path = hf_hub_download(
|
31 |
-
repo_id=repo_id,
|
32 |
-
filename=filename,
|
33 |
-
cache_dir=os.environ.get("HF_HOME", "./models")
|
34 |
-
)
|
35 |
-
|
36 |
-
# Load the checkpoint
|
37 |
-
return torch.load(local_path, map_location=device)
|
38 |
-
|
39 |
-
# Check if it's a HuggingFace repo format (username/repo/file.pt)
|
40 |
-
elif "/" in checkpoint_path and not os.path.exists(checkpoint_path):
|
41 |
-
parts = checkpoint_path.split("/")
|
42 |
-
if len(parts) >= 3:
|
43 |
-
repo_id = "/".join(parts[:2])
|
44 |
-
filename = "/".join(parts[2:])
|
45 |
-
|
46 |
-
print(f"Loading from HuggingFace: {repo_id}/{filename}")
|
47 |
-
|
48 |
-
local_path = hf_hub_download(
|
49 |
-
repo_id=repo_id,
|
50 |
-
filename=filename,
|
51 |
-
cache_dir=os.environ.get("HF_HOME", "./models")
|
52 |
-
)
|
53 |
-
|
54 |
-
return torch.load(local_path, map_location=device)
|
55 |
|
56 |
-
#
|
57 |
-
|
58 |
-
|
59 |
-
|
|
|
|
|
|
|
60 |
|
61 |
-
|
|
|
|
|
|
|
62 |
|
63 |
-
#
|
64 |
-
|
65 |
-
def __init__(
|
66 |
-
self,
|
67 |
-
student_checkpoint_path="",
|
68 |
-
duration_predictor_path="",
|
69 |
-
device="cuda",
|
70 |
-
model_type="F5TTS_Base",
|
71 |
-
tokenizer="pinyin",
|
72 |
-
dataset_name="Emilia_ZH_EN",
|
73 |
-
cuda_device_id="0"
|
74 |
-
):
|
75 |
-
# ... (previous initialization code) ...
|
76 |
-
|
77 |
-
# Initialize components
|
78 |
-
self._setup_tokenizer()
|
79 |
-
self._setup_models(student_checkpoint_path) # Modified to handle HF URLs
|
80 |
-
self._setup_mel_spec()
|
81 |
-
self._setup_vocoder()
|
82 |
-
self._setup_duration_predictor(duration_predictor_path) # Modified to handle HF URLs
|
83 |
-
|
84 |
-
def _setup_models(self, student_checkpoint_path):
|
85 |
-
"""Initialize teacher and student models with HF support."""
|
86 |
-
# ... (model configuration code) ...
|
87 |
-
|
88 |
-
# Load student checkpoint with HF support
|
89 |
-
checkpoint = load_checkpoint_from_hf(student_checkpoint_path, device='cpu')
|
90 |
-
self.model.load_state_dict(checkpoint['model_state_dict'], strict=False)
|
91 |
-
|
92 |
-
# ... (rest of the setup) ...
|
93 |
-
|
94 |
-
def _setup_duration_predictor(self, checkpoint_path):
|
95 |
-
"""Initialize duration predictor with HF support."""
|
96 |
-
# ... (model initialization code) ...
|
97 |
-
|
98 |
-
# Load checkpoint with HF support
|
99 |
-
checkpoint = load_checkpoint_from_hf(checkpoint_path, device='cpu')
|
100 |
-
self.SLP.load_state_dict(checkpoint['model_state_dict'])
|
101 |
|
102 |
-
|
103 |
-
|
104 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
105 |
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
if
|
111 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
|
113 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
114 |
|
115 |
-
|
116 |
-
"""Load checkpoint with HF URL support."""
|
117 |
-
return load_checkpoint_from_hf(checkpoint_path, self.device)
|
|
|
1 |
+
import gradio as gr
|
|
|
|
|
2 |
import torch
|
3 |
+
import torchaudio
|
4 |
+
import numpy as np
|
5 |
+
import tempfile
|
6 |
+
import time
|
7 |
from pathlib import Path
|
8 |
from huggingface_hub import hf_hub_download
|
9 |
+
import os
|
10 |
+
|
11 |
+
# Import the inference module (assuming it's named 'infer.py' based on the notebook)
|
12 |
+
from infer import DMOInference
|
13 |
+
|
14 |
+
# Global model instance
|
15 |
+
model = None
|
16 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
17 |
+
|
18 |
+
def download_models():
|
19 |
+
"""Download models from HuggingFace Hub."""
|
20 |
+
try:
|
21 |
+
print("Downloading models from HuggingFace...")
|
22 |
+
|
23 |
+
# Download student model
|
24 |
+
student_path = hf_hub_download(
|
25 |
+
repo_id="yl4579/DMOSpeech2",
|
26 |
+
filename="model_85000.pt",
|
27 |
+
cache_dir="./models"
|
28 |
+
)
|
29 |
+
|
30 |
+
# Download duration predictor
|
31 |
+
duration_path = hf_hub_download(
|
32 |
+
repo_id="yl4579/DMOSpeech2",
|
33 |
+
filename="model_1500.pt",
|
34 |
+
cache_dir="./models"
|
35 |
+
)
|
36 |
+
|
37 |
+
print(f"Student model: {student_path}")
|
38 |
+
print(f"Duration model: {duration_path}")
|
39 |
+
|
40 |
+
return student_path, duration_path
|
41 |
+
|
42 |
+
except Exception as e:
|
43 |
+
print(f"Error downloading models: {e}")
|
44 |
+
return None, None
|
45 |
|
46 |
+
def initialize_model():
|
47 |
+
"""Initialize the model on startup."""
|
48 |
+
global model
|
49 |
|
50 |
+
try:
|
51 |
+
# Download models
|
52 |
+
student_path, duration_path = download_models()
|
53 |
+
|
54 |
+
if not student_path or not duration_path:
|
55 |
+
return False, "Failed to download models from HuggingFace"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
56 |
|
57 |
+
# Initialize model
|
58 |
+
model = DMOInference(
|
59 |
+
student_checkpoint_path=student_path,
|
60 |
+
duration_predictor_path=duration_path,
|
61 |
+
device=device,
|
62 |
+
model_type="F5TTS_Base"
|
63 |
+
)
|
64 |
|
65 |
+
return True, f"Model loaded successfully on {device.upper()}"
|
66 |
+
|
67 |
+
except Exception as e:
|
68 |
+
return False, f"Error initializing model: {str(e)}"
|
69 |
|
70 |
+
# Initialize model on startup
|
71 |
+
model_loaded, status_message = initialize_model()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
|
73 |
+
def generate_speech(
|
74 |
+
prompt_audio,
|
75 |
+
prompt_text,
|
76 |
+
target_text,
|
77 |
+
mode,
|
78 |
+
# Advanced settings
|
79 |
+
custom_teacher_steps,
|
80 |
+
custom_teacher_stopping_time,
|
81 |
+
custom_student_start_step,
|
82 |
+
temperature,
|
83 |
+
verbose
|
84 |
+
):
|
85 |
+
"""Generate speech with different configurations."""
|
86 |
+
|
87 |
+
if not model_loaded or model is None:
|
88 |
+
return None, "Model not loaded! Please refresh the page.", "", ""
|
89 |
+
|
90 |
+
if prompt_audio is None:
|
91 |
+
return None, "Please upload a reference audio!", "", ""
|
92 |
+
|
93 |
+
if not target_text:
|
94 |
+
return None, "Please enter text to generate!", "", ""
|
95 |
|
96 |
+
try:
|
97 |
+
start_time = time.time()
|
98 |
+
|
99 |
+
# Configure parameters based on mode
|
100 |
+
if mode == "Student Only (4 steps)":
|
101 |
+
teacher_steps = 0
|
102 |
+
student_start_step = 0
|
103 |
+
teacher_stopping_time = 1.0
|
104 |
+
elif mode == "Teacher-Guided (8 steps)":
|
105 |
+
# Default configuration from the notebook
|
106 |
+
teacher_steps = 16
|
107 |
+
teacher_stopping_time = 0.07
|
108 |
+
student_start_step = 1
|
109 |
+
elif mode == "High Diversity (16 steps)":
|
110 |
+
teacher_steps = 24
|
111 |
+
teacher_stopping_time = 0.3
|
112 |
+
student_start_step = 2
|
113 |
+
else: # Custom
|
114 |
+
teacher_steps = custom_teacher_steps
|
115 |
+
teacher_stopping_time = custom_teacher_stopping_time
|
116 |
+
student_start_step = custom_student_start_step
|
117 |
+
|
118 |
+
# Generate speech
|
119 |
+
generated_audio = model.generate(
|
120 |
+
gen_text=target_text,
|
121 |
+
audio_path=prompt_audio,
|
122 |
+
prompt_text=prompt_text if prompt_text else None,
|
123 |
+
teacher_steps=teacher_steps,
|
124 |
+
teacher_stopping_time=teacher_stopping_time,
|
125 |
+
student_start_step=student_start_step,
|
126 |
+
temperature=temperature,
|
127 |
+
verbose=verbose
|
128 |
+
)
|
129 |
+
|
130 |
+
end_time = time.time()
|
131 |
+
|
132 |
+
# Calculate metrics
|
133 |
+
processing_time = end_time - start_time
|
134 |
+
audio_duration = generated_audio.shape[-1] / 24000
|
135 |
+
rtf = processing_time / audio_duration
|
136 |
+
|
137 |
+
# Save audio
|
138 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_file:
|
139 |
+
output_path = tmp_file.name
|
140 |
+
|
141 |
+
if isinstance(generated_audio, np.ndarray):
|
142 |
+
generated_audio = torch.from_numpy(generated_audio)
|
143 |
+
|
144 |
+
if generated_audio.dim() == 1:
|
145 |
+
generated_audio = generated_audio.unsqueeze(0)
|
146 |
|
147 |
+
torchaudio.save(output_path, generated_audio, 24000)
|
148 |
+
|
149 |
+
# Format metrics
|
150 |
+
metrics = f"RTF: {rtf:.2f}x ({1/rtf:.2f}x speed) | Processing: {processing_time:.2f}s for {audio_duration:.2f}s audio"
|
151 |
+
|
152 |
+
return output_path, "Success!", metrics, f"Mode: {mode}"
|
153 |
+
|
154 |
+
except Exception as e:
|
155 |
+
return None, f"Error: {str(e)}", "", ""
|
156 |
+
|
157 |
+
# Create Gradio interface
|
158 |
+
with gr.Blocks(title="DMOSpeech 2 - Zero-Shot TTS", theme=gr.themes.Soft()) as demo:
|
159 |
+
gr.Markdown(f"""
|
160 |
+
# 🎙️ DMOSpeech 2: Zero-Shot Text-to-Speech
|
161 |
+
|
162 |
+
Generate natural speech in any voice with just a short reference audio!
|
163 |
+
|
164 |
+
**Model Status:** {status_message} | **Device:** {device.upper()}
|
165 |
+
""")
|
166 |
+
|
167 |
+
with gr.Row():
|
168 |
+
with gr.Column(scale=1):
|
169 |
+
# Reference audio input
|
170 |
+
prompt_audio = gr.Audio(
|
171 |
+
label="📎 Reference Audio",
|
172 |
+
type="filepath",
|
173 |
+
sources=["upload", "microphone"]
|
174 |
+
)
|
175 |
+
|
176 |
+
prompt_text = gr.Textbox(
|
177 |
+
label="📝 Reference Text (optional - will auto-transcribe if empty)",
|
178 |
+
placeholder="The text spoken in the reference audio...",
|
179 |
+
lines=2
|
180 |
+
)
|
181 |
+
|
182 |
+
target_text = gr.Textbox(
|
183 |
+
label="✍️ Text to Generate",
|
184 |
+
placeholder="Enter the text you want to synthesize...",
|
185 |
+
lines=4
|
186 |
+
)
|
187 |
+
|
188 |
+
# Generation mode
|
189 |
+
mode = gr.Radio(
|
190 |
+
choices=[
|
191 |
+
"Student Only (4 steps)",
|
192 |
+
"Teacher-Guided (8 steps)",
|
193 |
+
"High Diversity (16 steps)",
|
194 |
+
"Custom"
|
195 |
+
],
|
196 |
+
value="Teacher-Guided (8 steps)",
|
197 |
+
label="🚀 Generation Mode",
|
198 |
+
info="Choose speed vs quality/diversity tradeoff"
|
199 |
+
)
|
200 |
+
|
201 |
+
# Advanced settings (collapsible)
|
202 |
+
with gr.Accordion("⚙️ Advanced Settings", open=False):
|
203 |
+
with gr.Row():
|
204 |
+
custom_teacher_steps = gr.Slider(
|
205 |
+
minimum=0,
|
206 |
+
maximum=32,
|
207 |
+
value=16,
|
208 |
+
step=1,
|
209 |
+
label="Teacher Steps",
|
210 |
+
info="More steps = higher quality"
|
211 |
+
)
|
212 |
+
|
213 |
+
custom_teacher_stopping_time = gr.Slider(
|
214 |
+
minimum=0.0,
|
215 |
+
maximum=1.0,
|
216 |
+
value=0.07,
|
217 |
+
step=0.01,
|
218 |
+
label="Teacher Stopping Time",
|
219 |
+
info="When to switch to student"
|
220 |
+
)
|
221 |
+
|
222 |
+
custom_student_start_step = gr.Slider(
|
223 |
+
minimum=0,
|
224 |
+
maximum=4,
|
225 |
+
value=1,
|
226 |
+
step=1,
|
227 |
+
label="Student Start Step",
|
228 |
+
info="Which student step to start from"
|
229 |
+
)
|
230 |
+
|
231 |
+
temperature = gr.Slider(
|
232 |
+
minimum=0.0,
|
233 |
+
maximum=2.0,
|
234 |
+
value=0.0,
|
235 |
+
step=0.1,
|
236 |
+
label="Duration Temperature",
|
237 |
+
info="0 = deterministic, >0 = more variation in speech rhythm"
|
238 |
+
)
|
239 |
+
|
240 |
+
verbose = gr.Checkbox(
|
241 |
+
value=False,
|
242 |
+
label="Verbose Output",
|
243 |
+
info="Show detailed generation steps"
|
244 |
+
)
|
245 |
+
|
246 |
+
generate_btn = gr.Button("🎵 Generate Speech", variant="primary", size="lg")
|
247 |
+
|
248 |
+
with gr.Column(scale=1):
|
249 |
+
# Output
|
250 |
+
output_audio = gr.Audio(
|
251 |
+
label="🔊 Generated Speech",
|
252 |
+
type="filepath",
|
253 |
+
autoplay=True
|
254 |
+
)
|
255 |
+
|
256 |
+
status = gr.Textbox(
|
257 |
+
label="Status",
|
258 |
+
interactive=False
|
259 |
+
)
|
260 |
+
|
261 |
+
metrics = gr.Textbox(
|
262 |
+
label="Performance Metrics",
|
263 |
+
interactive=False
|
264 |
+
)
|
265 |
+
|
266 |
+
info = gr.Textbox(
|
267 |
+
label="Generation Info",
|
268 |
+
interactive=False
|
269 |
+
)
|
270 |
+
|
271 |
+
# Tips
|
272 |
+
gr.Markdown("""
|
273 |
+
### 💡 Quick Tips:
|
274 |
+
|
275 |
+
- **Student Only**: Fastest (4 steps), good quality
|
276 |
+
- **Teacher-Guided**: Best balance (8 steps), recommended
|
277 |
+
- **High Diversity**: More natural prosody (16 steps)
|
278 |
+
- **Temperature**: Add randomness to speech rhythm
|
279 |
+
|
280 |
+
### 📊 Expected RTF (Real-Time Factor):
|
281 |
+
- Student Only: ~0.05x (20x faster than real-time)
|
282 |
+
- Teacher-Guided: ~0.10x (10x faster)
|
283 |
+
- High Diversity: ~0.20x (5x faster)
|
284 |
+
""")
|
285 |
+
|
286 |
+
# Examples section
|
287 |
+
gr.Markdown("### 🎯 Examples")
|
288 |
+
|
289 |
+
examples = [
|
290 |
+
[
|
291 |
+
None, # Will be replaced with actual audio path
|
292 |
+
"Some call me nature, others call me mother nature.",
|
293 |
+
"I don't really care what you call me. I've been a silent spectator, watching species evolve, empires rise and fall. But always remember, I am mighty and enduring.",
|
294 |
+
"Teacher-Guided (8 steps)",
|
295 |
+
16, 0.07, 1, 0.0, False
|
296 |
+
],
|
297 |
+
[
|
298 |
+
None, # Will be replaced with actual audio path
|
299 |
+
"对,这就是我,万人敬仰的太乙真人。",
|
300 |
+
'突然,身边一阵笑声。我看着他们,意气风发地挺直了胸膛,甩了甩那稍显肉感的双臂,轻笑道:"我身上的肉,是为了掩饰我爆棚的魅力,否则,岂不吓坏了你们呢?"',
|
301 |
+
"Teacher-Guided (8 steps)",
|
302 |
+
16, 0.07, 1, 0.0, False
|
303 |
+
],
|
304 |
+
[
|
305 |
+
None,
|
306 |
+
"对,这就是我,万人敬仰的太乙真人。",
|
307 |
+
'突然,身边一阵笑声。我看着他们,意气风发地挺直了胸膛,甩了甩那稍显肉感的双臂,轻笑道:"我身上��肉,是为了掩饰我爆棚的魅力,否则,岂不吓坏了你们呢?"',
|
308 |
+
"High Diversity (16 steps)",
|
309 |
+
24, 0.3, 2, 0.8, False
|
310 |
+
]
|
311 |
+
]
|
312 |
+
|
313 |
+
# Note about example audio files
|
314 |
+
gr.Markdown("""
|
315 |
+
*Note: Example audio files should be uploaded to the Space. The examples above show the text configurations used in the original notebook.*
|
316 |
+
""")
|
317 |
+
|
318 |
+
# Event handler
|
319 |
+
generate_btn.click(
|
320 |
+
generate_speech,
|
321 |
+
inputs=[
|
322 |
+
prompt_audio,
|
323 |
+
prompt_text,
|
324 |
+
target_text,
|
325 |
+
mode,
|
326 |
+
custom_teacher_steps,
|
327 |
+
custom_teacher_stopping_time,
|
328 |
+
custom_student_start_step,
|
329 |
+
temperature,
|
330 |
+
verbose
|
331 |
+
],
|
332 |
+
outputs=[output_audio, status, metrics, info]
|
333 |
+
)
|
334 |
+
|
335 |
+
# Update visibility of custom settings based on mode
|
336 |
+
def update_custom_visibility(mode):
|
337 |
+
return gr.update(visible=(mode == "Custom"))
|
338 |
+
|
339 |
+
mode.change(
|
340 |
+
lambda x: [gr.update(interactive=(x == "Custom"))] * 3,
|
341 |
+
inputs=[mode],
|
342 |
+
outputs=[custom_teacher_steps, custom_teacher_stopping_time, custom_student_start_step]
|
343 |
+
)
|
344 |
+
|
345 |
+
# Launch the app
|
346 |
+
if __name__ == "__main__":
|
347 |
+
if not model_loaded:
|
348 |
+
print(f"Warning: Model failed to load - {status_message}")
|
349 |
|
350 |
+
demo.launch()
|
|
|
|