Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -26,6 +26,10 @@ except ImportError as e:
|
|
26 |
model = None
|
27 |
model_loaded = False
|
28 |
|
|
|
|
|
|
|
|
|
29 |
def download_model_files():
|
30 |
"""Download model files with error handling."""
|
31 |
print(f"Checking for model files in {LOCAL_MODEL_PATH}...")
|
@@ -65,7 +69,6 @@ def load_model_on_gpu():
|
|
65 |
try:
|
66 |
print("Loading model inside GPU context...")
|
67 |
|
68 |
-
# Now we can safely use CUDA operations
|
69 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
70 |
print(f"Loading model on device: {device}")
|
71 |
|
@@ -80,10 +83,8 @@ def load_model_on_gpu():
|
|
80 |
print("β Model loaded successfully with from_pretrained.")
|
81 |
except Exception as e2:
|
82 |
print(f"from_pretrained failed: {e2}")
|
83 |
-
# Manual loading as fallback
|
84 |
model = load_model_manually(device)
|
85 |
|
86 |
-
# Move model to device and set to eval mode
|
87 |
if model and hasattr(model, 'to'):
|
88 |
model = model.to(device)
|
89 |
if model and hasattr(model, 'eval'):
|
@@ -108,7 +109,6 @@ def load_model_manually(device):
|
|
108 |
model_path = pathlib.Path(LOCAL_MODEL_PATH)
|
109 |
print("Manual loading with correct constructor signature...")
|
110 |
|
111 |
-
# Load components to CPU first, then move to device
|
112 |
s3gen_path = model_path / "s3gen.pt"
|
113 |
ve_path = model_path / "ve.pt"
|
114 |
tokenizer_path = model_path / "tokenizer.json"
|
@@ -127,7 +127,6 @@ def load_model_manually(device):
|
|
127 |
except Exception:
|
128 |
tokenizer = tokenizer_data
|
129 |
|
130 |
-
# Create model instance
|
131 |
model = ChatterboxTTS(
|
132 |
t3=t3_cfg,
|
133 |
s3gen=s3gen,
|
@@ -141,10 +140,35 @@ def load_model_manually(device):
|
|
141 |
|
142 |
def cleanup_gpu_memory():
|
143 |
"""Clean up GPU memory - only call within GPU context."""
|
144 |
-
|
145 |
-
torch.cuda.
|
146 |
-
|
147 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
148 |
|
149 |
# Download model files during startup (CPU only)
|
150 |
if chatterbox_available:
|
@@ -169,8 +193,16 @@ def clone_voice(text_to_speak, reference_audio_path, exaggeration=0.6, cfg_pace=
|
|
169 |
if reference_audio_path is None:
|
170 |
return None, "Error: Please upload a reference audio file (.wav or .mp3)."
|
171 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
172 |
try:
|
173 |
-
# Load model if not already loaded
|
174 |
if not model_loaded:
|
175 |
print("Loading model for the first time...")
|
176 |
if not load_model_on_gpu():
|
@@ -180,7 +212,9 @@ def clone_voice(text_to_speak, reference_audio_path, exaggeration=0.6, cfg_pace=
|
|
180 |
return None, "Error: Model not loaded. Please check the logs for details."
|
181 |
|
182 |
print(f"Processing request:")
|
183 |
-
print(f"
|
|
|
|
|
184 |
print(f" Audio: '{reference_audio_path}'")
|
185 |
print(f" Parameters: exag={exaggeration}, cfg={cfg_pace}, seed={random_seed}, temp={temperature}")
|
186 |
|
@@ -199,33 +233,19 @@ def clone_voice(text_to_speak, reference_audio_path, exaggeration=0.6, cfg_pace=
|
|
199 |
|
200 |
# Generate audio with error handling
|
201 |
try:
|
202 |
-
with torch.no_grad():
|
203 |
output_wav_data = model.generate(
|
204 |
-
text=
|
205 |
audio_prompt_path=reference_audio_path,
|
206 |
exaggeration=exaggeration,
|
207 |
cfg_weight=cfg_pace,
|
208 |
temperature=temperature
|
209 |
)
|
210 |
except RuntimeError as e:
|
211 |
-
if "CUDA" in str(e) or "out of memory" in str(e):
|
212 |
print(f"CUDA error during generation: {e}")
|
213 |
-
# Try to recover by cleaning memory and retrying
|
214 |
cleanup_gpu_memory()
|
215 |
-
|
216 |
-
with torch.no_grad():
|
217 |
-
output_wav_data = model.generate(
|
218 |
-
text=text_to_speak,
|
219 |
-
audio_prompt_path=reference_audio_path,
|
220 |
-
exaggeration=exaggeration,
|
221 |
-
cfg_weight=cfg_pace,
|
222 |
-
temperature=temperature
|
223 |
-
)
|
224 |
-
print("β Recovery successful after memory cleanup")
|
225 |
-
except Exception as retry_error:
|
226 |
-
print(f"β Recovery failed: {retry_error}")
|
227 |
-
cleanup_gpu_memory()
|
228 |
-
return None, f"CUDA error: {str(e)}. GPU memory issue - please try again in a moment."
|
229 |
else:
|
230 |
raise e
|
231 |
|
@@ -253,7 +273,13 @@ def clone_voice(text_to_speak, reference_audio_path, exaggeration=0.6, cfg_pace=
|
|
253 |
print(f"CUDA memory after generation: {torch.cuda.memory_allocated() / 1024**2:.1f} MB")
|
254 |
|
255 |
print("β Audio generated successfully")
|
256 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
257 |
|
258 |
except Exception as e:
|
259 |
print(f"ERROR during audio generation: {e}")
|
@@ -268,14 +294,14 @@ def clone_voice(text_to_speak, reference_audio_path, exaggeration=0.6, cfg_pace=
|
|
268 |
# Provide specific error messages
|
269 |
error_msg = str(e)
|
270 |
if "CUDA" in error_msg or "device-side assert" in error_msg:
|
271 |
-
return None, f"CUDA error: {error_msg}.
|
272 |
elif "out of memory" in error_msg:
|
273 |
-
return None, f"GPU memory error: {error_msg}. Please try with shorter text
|
274 |
else:
|
275 |
return None, f"Error during audio generation: {error_msg}. Check logs for more details."
|
276 |
|
277 |
def clone_voice_api(text_to_speak, reference_audio_url, exaggeration=0.6, cfg_pace=0.3, random_seed=0, temperature=0.6):
|
278 |
-
"""API wrapper function
|
279 |
import requests
|
280 |
import tempfile
|
281 |
import os
|
@@ -317,21 +343,28 @@ def clone_voice_api(text_to_speak, reference_audio_url, exaggeration=0.6, cfg_pa
|
|
317 |
except:
|
318 |
pass
|
319 |
|
320 |
-
# Your existing Gradio interface code goes here...
|
321 |
def main():
|
322 |
print("Starting Advanced Gradio interface...")
|
323 |
|
324 |
-
# Your existing Gradio interface code
|
325 |
with gr.Blocks(title="ποΈ Advanced Chatterbox Voice Cloning") as demo:
|
326 |
gr.Markdown("# ποΈ Advanced Chatterbox Voice Cloning")
|
327 |
gr.Markdown("Clone any voice using advanced AI technology with fine-tuned controls.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
328 |
|
329 |
with gr.Row():
|
330 |
with gr.Column(scale=2):
|
331 |
text_input = gr.Textbox(
|
332 |
-
label="Text to Speak",
|
333 |
placeholder="Enter the text you want the cloned voice to say...",
|
334 |
-
lines=
|
|
|
335 |
)
|
336 |
audio_input = gr.Audio(
|
337 |
type="filepath",
|
@@ -362,7 +395,7 @@ def main():
|
|
362 |
|
363 |
with gr.Column(scale=1):
|
364 |
audio_output = gr.Audio(label="Generated Audio", type="numpy")
|
365 |
-
status_output = gr.Textbox(label="Status", lines=
|
366 |
|
367 |
# Connect the interface
|
368 |
generate_btn.click(
|
|
|
26 |
model = None
|
27 |
model_loaded = False
|
28 |
|
29 |
+
# Text length limits for the model
|
30 |
+
MAX_CHARS_PER_GENERATION = 1000 # Safe limit for single generation
|
31 |
+
MAX_CHARS_TOTAL = 5000 # Maximum we'll accept via API
|
32 |
+
|
33 |
def download_model_files():
|
34 |
"""Download model files with error handling."""
|
35 |
print(f"Checking for model files in {LOCAL_MODEL_PATH}...")
|
|
|
69 |
try:
|
70 |
print("Loading model inside GPU context...")
|
71 |
|
|
|
72 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
73 |
print(f"Loading model on device: {device}")
|
74 |
|
|
|
83 |
print("β Model loaded successfully with from_pretrained.")
|
84 |
except Exception as e2:
|
85 |
print(f"from_pretrained failed: {e2}")
|
|
|
86 |
model = load_model_manually(device)
|
87 |
|
|
|
88 |
if model and hasattr(model, 'to'):
|
89 |
model = model.to(device)
|
90 |
if model and hasattr(model, 'eval'):
|
|
|
109 |
model_path = pathlib.Path(LOCAL_MODEL_PATH)
|
110 |
print("Manual loading with correct constructor signature...")
|
111 |
|
|
|
112 |
s3gen_path = model_path / "s3gen.pt"
|
113 |
ve_path = model_path / "ve.pt"
|
114 |
tokenizer_path = model_path / "tokenizer.json"
|
|
|
127 |
except Exception:
|
128 |
tokenizer = tokenizer_data
|
129 |
|
|
|
130 |
model = ChatterboxTTS(
|
131 |
t3=t3_cfg,
|
132 |
s3gen=s3gen,
|
|
|
140 |
|
141 |
def cleanup_gpu_memory():
|
142 |
"""Clean up GPU memory - only call within GPU context."""
|
143 |
+
try:
|
144 |
+
if torch.cuda.is_available():
|
145 |
+
torch.cuda.empty_cache()
|
146 |
+
torch.cuda.synchronize()
|
147 |
+
gc.collect()
|
148 |
+
except Exception as e:
|
149 |
+
print(f"Warning: GPU cleanup failed: {e}")
|
150 |
+
|
151 |
+
def truncate_text_safely(text, max_chars=MAX_CHARS_PER_GENERATION):
|
152 |
+
"""Truncate text to safe length while preserving sentence boundaries."""
|
153 |
+
if len(text) <= max_chars:
|
154 |
+
return text, False
|
155 |
+
|
156 |
+
# Find the last sentence ending before the limit
|
157 |
+
truncated = text[:max_chars]
|
158 |
+
|
159 |
+
# Look for sentence endings
|
160 |
+
for ending in ['. ', '! ', '? ']:
|
161 |
+
last_sentence = truncated.rfind(ending)
|
162 |
+
if last_sentence > max_chars * 0.7: # Don't truncate too aggressively
|
163 |
+
return text[:last_sentence + 1].strip(), True
|
164 |
+
|
165 |
+
# Fallback to word boundary
|
166 |
+
last_space = truncated.rfind(' ')
|
167 |
+
if last_space > max_chars * 0.8:
|
168 |
+
return text[:last_space].strip(), True
|
169 |
+
|
170 |
+
# Last resort: hard truncate
|
171 |
+
return truncated.strip(), True
|
172 |
|
173 |
# Download model files during startup (CPU only)
|
174 |
if chatterbox_available:
|
|
|
193 |
if reference_audio_path is None:
|
194 |
return None, "Error: Please upload a reference audio file (.wav or .mp3)."
|
195 |
|
196 |
+
# Check text length and truncate if necessary
|
197 |
+
original_length = len(text_to_speak)
|
198 |
+
if original_length > MAX_CHARS_TOTAL:
|
199 |
+
return None, f"Error: Text is too long ({original_length:,} characters). Maximum allowed is {MAX_CHARS_TOTAL:,} characters. Please use the chunked generation API for longer texts."
|
200 |
+
|
201 |
+
# Truncate to safe generation length
|
202 |
+
text_to_use, was_truncated = truncate_text_safely(text_to_speak, MAX_CHARS_PER_GENERATION)
|
203 |
+
|
204 |
try:
|
205 |
+
# Load model if not already loaded
|
206 |
if not model_loaded:
|
207 |
print("Loading model for the first time...")
|
208 |
if not load_model_on_gpu():
|
|
|
212 |
return None, "Error: Model not loaded. Please check the logs for details."
|
213 |
|
214 |
print(f"Processing request:")
|
215 |
+
print(f" Original text length: {original_length:,} characters")
|
216 |
+
print(f" Processing length: {len(text_to_use):,} characters")
|
217 |
+
print(f" Truncated: {was_truncated}")
|
218 |
print(f" Audio: '{reference_audio_path}'")
|
219 |
print(f" Parameters: exag={exaggeration}, cfg={cfg_pace}, seed={random_seed}, temp={temperature}")
|
220 |
|
|
|
233 |
|
234 |
# Generate audio with error handling
|
235 |
try:
|
236 |
+
with torch.no_grad():
|
237 |
output_wav_data = model.generate(
|
238 |
+
text=text_to_use,
|
239 |
audio_prompt_path=reference_audio_path,
|
240 |
exaggeration=exaggeration,
|
241 |
cfg_weight=cfg_pace,
|
242 |
temperature=temperature
|
243 |
)
|
244 |
except RuntimeError as e:
|
245 |
+
if "CUDA" in str(e) or "out of memory" in str(e) or "device-side assert" in str(e):
|
246 |
print(f"CUDA error during generation: {e}")
|
|
|
247 |
cleanup_gpu_memory()
|
248 |
+
return None, f"CUDA error: Text may be too long for single generation. Try shorter text (under {MAX_CHARS_PER_GENERATION} characters) or use the chunked generation API for longer content."
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
249 |
else:
|
250 |
raise e
|
251 |
|
|
|
273 |
print(f"CUDA memory after generation: {torch.cuda.memory_allocated() / 1024**2:.1f} MB")
|
274 |
|
275 |
print("β Audio generated successfully")
|
276 |
+
|
277 |
+
# Prepare success message
|
278 |
+
success_msg = "Success: Audio generated successfully!"
|
279 |
+
if was_truncated:
|
280 |
+
success_msg += f" Note: Text was truncated from {original_length:,} to {len(text_to_use):,} characters for optimal generation. Use the chunked generation API for longer texts."
|
281 |
+
|
282 |
+
return result, success_msg
|
283 |
|
284 |
except Exception as e:
|
285 |
print(f"ERROR during audio generation: {e}")
|
|
|
294 |
# Provide specific error messages
|
295 |
error_msg = str(e)
|
296 |
if "CUDA" in error_msg or "device-side assert" in error_msg:
|
297 |
+
return None, f"CUDA error: {error_msg}. Try shorter text (under {MAX_CHARS_PER_GENERATION} characters) or use the chunked generation API."
|
298 |
elif "out of memory" in error_msg:
|
299 |
+
return None, f"GPU memory error: {error_msg}. Please try with shorter text."
|
300 |
else:
|
301 |
return None, f"Error during audio generation: {error_msg}. Check logs for more details."
|
302 |
|
303 |
def clone_voice_api(text_to_speak, reference_audio_url, exaggeration=0.6, cfg_pace=0.3, random_seed=0, temperature=0.6):
|
304 |
+
"""API wrapper function."""
|
305 |
import requests
|
306 |
import tempfile
|
307 |
import os
|
|
|
343 |
except:
|
344 |
pass
|
345 |
|
|
|
346 |
def main():
|
347 |
print("Starting Advanced Gradio interface...")
|
348 |
|
|
|
349 |
with gr.Blocks(title="ποΈ Advanced Chatterbox Voice Cloning") as demo:
|
350 |
gr.Markdown("# ποΈ Advanced Chatterbox Voice Cloning")
|
351 |
gr.Markdown("Clone any voice using advanced AI technology with fine-tuned controls.")
|
352 |
+
|
353 |
+
# Add warning about text length
|
354 |
+
gr.Markdown(f"""
|
355 |
+
**β οΈ Text Length Limits:**
|
356 |
+
- **Single Generation**: Up to {MAX_CHARS_PER_GENERATION:,} characters (optimal quality)
|
357 |
+
- **API Maximum**: Up to {MAX_CHARS_TOTAL:,} characters (may be truncated)
|
358 |
+
- **For longer texts**: Use the chunked generation API in your application
|
359 |
+
""")
|
360 |
|
361 |
with gr.Row():
|
362 |
with gr.Column(scale=2):
|
363 |
text_input = gr.Textbox(
|
364 |
+
label=f"Text to Speak (max {MAX_CHARS_TOTAL:,} characters)",
|
365 |
placeholder="Enter the text you want the cloned voice to say...",
|
366 |
+
lines=5,
|
367 |
+
max_lines=10
|
368 |
)
|
369 |
audio_input = gr.Audio(
|
370 |
type="filepath",
|
|
|
395 |
|
396 |
with gr.Column(scale=1):
|
397 |
audio_output = gr.Audio(label="Generated Audio", type="numpy")
|
398 |
+
status_output = gr.Textbox(label="Status", lines=3)
|
399 |
|
400 |
# Connect the interface
|
401 |
generate_btn.click(
|