Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,9 +1,7 @@
|
|
1 |
import os
|
2 |
import re
|
3 |
-
import logging
|
4 |
import torch
|
5 |
import tempfile
|
6 |
-
from typing import Tuple, Union
|
7 |
from scipy.io.wavfile import write
|
8 |
from pydub import AudioSegment
|
9 |
from dotenv import load_dotenv
|
@@ -21,149 +19,102 @@ from transformers import (
|
|
21 |
# Coqui TTS
|
22 |
from TTS.api import TTS
|
23 |
|
24 |
-
# Kokoro TTS (ensure these are installed)
|
25 |
-
# pip install -q kokoro>=0.8.2 soundfile
|
26 |
-
# apt-get -qq -y install espeak-ng > /dev/null 2>&1
|
27 |
-
from kokoro import KPipeline
|
28 |
-
import soundfile as sf
|
29 |
-
|
30 |
# ---------------------------------------------------------------------
|
31 |
-
#
|
32 |
# ---------------------------------------------------------------------
|
33 |
load_dotenv()
|
34 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
35 |
-
if not HF_TOKEN:
|
36 |
-
logging.warning("HF_TOKEN environment variable not set!")
|
37 |
-
|
38 |
-
logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
|
39 |
|
|
|
40 |
# Global Model Caches
|
|
|
41 |
LLAMA_PIPELINES = {}
|
42 |
MUSICGEN_MODELS = {}
|
43 |
TTS_MODELS = {}
|
44 |
|
45 |
# ---------------------------------------------------------------------
|
46 |
-
# Utility
|
47 |
# ---------------------------------------------------------------------
|
48 |
def clean_text(text: str) -> str:
|
49 |
"""
|
50 |
-
|
51 |
-
|
52 |
-
Args:
|
53 |
-
text (str): Input text to be cleaned.
|
54 |
-
|
55 |
-
Returns:
|
56 |
-
str: Cleaned text.
|
57 |
"""
|
58 |
-
# Remove all asterisks.
|
59 |
return re.sub(r'\*', '', text)
|
60 |
|
61 |
# ---------------------------------------------------------------------
|
62 |
-
#
|
63 |
# ---------------------------------------------------------------------
|
64 |
-
def get_llama_pipeline(model_id: str, token: str)
|
65 |
"""
|
66 |
-
|
67 |
-
|
68 |
-
Args:
|
69 |
-
model_id (str): Hugging Face model identifier.
|
70 |
-
token (str): Hugging Face authentication token.
|
71 |
-
|
72 |
-
Returns:
|
73 |
-
pipeline: Text-generation pipeline instance.
|
74 |
"""
|
75 |
if model_id in LLAMA_PIPELINES:
|
76 |
return LLAMA_PIPELINES[model_id]
|
77 |
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
return text_pipeline
|
90 |
-
except Exception as e:
|
91 |
-
logging.error(f"Error loading LLaMA pipeline: {e}")
|
92 |
-
raise
|
93 |
|
94 |
-
|
|
|
95 |
"""
|
96 |
-
|
97 |
-
|
98 |
-
Args:
|
99 |
-
model_key (str): Model key (default uses 'facebook/musicgen-large').
|
100 |
-
|
101 |
-
Returns:
|
102 |
-
tuple: (MusicGen model, processor)
|
103 |
"""
|
104 |
if model_key in MUSICGEN_MODELS:
|
105 |
return MUSICGEN_MODELS[model_key]
|
106 |
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
except Exception as e:
|
115 |
-
logging.error(f"Error loading MusicGen model: {e}")
|
116 |
-
raise
|
117 |
|
118 |
-
|
|
|
119 |
"""
|
120 |
-
|
121 |
-
|
122 |
-
Args:
|
123 |
-
model_name (str): Name of the TTS model.
|
124 |
-
|
125 |
-
Returns:
|
126 |
-
TTS: TTS model instance.
|
127 |
"""
|
128 |
if model_name in TTS_MODELS:
|
129 |
return TTS_MODELS[model_name]
|
130 |
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
except Exception as e:
|
136 |
-
logging.error(f"Error loading TTS model: {e}")
|
137 |
-
raise
|
138 |
|
139 |
# ---------------------------------------------------------------------
|
140 |
# Script Generation Function
|
141 |
# ---------------------------------------------------------------------
|
142 |
@spaces.GPU(duration=100)
|
143 |
-
def generate_script(user_prompt: str, model_id: str, token: str, duration: int)
|
144 |
"""
|
145 |
-
|
146 |
-
|
147 |
-
Args:
|
148 |
-
user_prompt (str): The user's creative input.
|
149 |
-
model_id (str): Hugging Face model identifier for LLaMA.
|
150 |
-
token (str): Hugging Face authentication token.
|
151 |
-
duration (int): Desired duration of the promo in seconds.
|
152 |
-
|
153 |
-
Returns:
|
154 |
-
tuple: (voice_script, sound_design, music_suggestions)
|
155 |
"""
|
156 |
try:
|
157 |
text_pipeline = get_llama_pipeline(model_id, token)
|
|
|
158 |
system_prompt = (
|
159 |
"You are an expert radio imaging producer specializing in sound design and music. "
|
160 |
-
f"Based on the user's concept and the selected duration of {duration} seconds, produce the following
|
161 |
-
"1. A concise voice-over script. Prefix this section with 'Voice-Over Script:'
|
162 |
-
"2. Suggestions for sound design. Prefix this section with 'Sound Design Suggestions:'
|
163 |
-
"3. Music styles or track recommendations. Prefix this section with 'Music Suggestions:'"
|
164 |
)
|
165 |
combined_prompt = f"{system_prompt}\nUser concept: {user_prompt}\nOutput:"
|
166 |
-
|
167 |
with torch.inference_mode():
|
168 |
result = text_pipeline(
|
169 |
combined_prompt,
|
@@ -173,130 +124,86 @@ def generate_script(user_prompt: str, model_id: str, token: str, duration: int)
|
|
173 |
)
|
174 |
|
175 |
generated_text = result[0]["generated_text"]
|
176 |
-
# Remove everything before the 'Output:' marker if present
|
177 |
if "Output:" in generated_text:
|
178 |
generated_text = generated_text.split("Output:")[-1].strip()
|
179 |
|
180 |
-
#
|
181 |
voice_script = "No voice-over script found."
|
182 |
sound_design = "No sound design suggestions found."
|
183 |
music_suggestions = "No music suggestions found."
|
184 |
|
185 |
-
#
|
186 |
if "Voice-Over Script:" in generated_text:
|
187 |
-
|
188 |
-
|
189 |
-
|
|
|
190 |
else:
|
191 |
-
voice_script =
|
192 |
|
|
|
193 |
if "Sound Design Suggestions:" in generated_text:
|
194 |
-
|
195 |
-
|
196 |
-
|
|
|
197 |
else:
|
198 |
-
sound_design =
|
199 |
|
|
|
200 |
if "Music Suggestions:" in generated_text:
|
201 |
-
|
|
|
202 |
|
203 |
return voice_script, sound_design, music_suggestions
|
204 |
|
205 |
except Exception as e:
|
206 |
-
logging.error(f"Error in generate_script: {e}")
|
207 |
return f"Error generating script: {e}", "", ""
|
208 |
|
|
|
209 |
# ---------------------------------------------------------------------
|
210 |
-
# Voice-Over Generation
|
211 |
# ---------------------------------------------------------------------
|
212 |
@spaces.GPU(duration=100)
|
213 |
-
def generate_voice(script: str, tts_model_name: str = "tts_models/en/ljspeech/tacotron2-DDC")
|
214 |
"""
|
215 |
-
|
216 |
-
|
217 |
-
Args:
|
218 |
-
script (str): The voice-over script.
|
219 |
-
tts_model_name (str): TTS model identifier.
|
220 |
-
|
221 |
-
Returns:
|
222 |
-
str: File path to the generated .wav file or an error message.
|
223 |
"""
|
224 |
try:
|
225 |
if not script.strip():
|
226 |
-
|
|
|
|
|
227 |
cleaned_script = clean_text(script)
|
|
|
228 |
tts_model = get_tts_model(tts_model_name)
|
229 |
-
|
|
|
|
|
230 |
tts_model.tts_to_file(text=cleaned_script, file_path=output_path)
|
231 |
-
logging.info(f"Coqui voice-over generated at {output_path}")
|
232 |
return output_path
|
233 |
|
234 |
except Exception as e:
|
235 |
-
logging.error(f"Error in generate_voice (Coqui TTS): {e}")
|
236 |
return f"Error generating voice: {e}"
|
237 |
|
238 |
-
@spaces.GPU(duration=100)
|
239 |
-
def generate_voice_kokoro(script: str, lang_code: str = 'a', voice: str = 'af_heart', speed: float = 1.0) -> Union[str, None]:
|
240 |
-
"""
|
241 |
-
Generate a voice-over audio file using the Kokoro TTS model.
|
242 |
-
|
243 |
-
Args:
|
244 |
-
script (str): The text to synthesize.
|
245 |
-
lang_code (str): Language code ('a' for American English, etc.).
|
246 |
-
voice (str): Specific voice style.
|
247 |
-
speed (float): Speech speed.
|
248 |
-
|
249 |
-
Returns:
|
250 |
-
str: File path to the generated WAV file or an error message.
|
251 |
-
"""
|
252 |
-
try:
|
253 |
-
# Initialize the Kokoro pipeline
|
254 |
-
kp = KPipeline(lang_code=lang_code)
|
255 |
-
audio_segments = []
|
256 |
-
generator = kp(script, voice=voice, speed=speed, split_pattern=r'\n+')
|
257 |
-
for i, (gs, ps, audio) in enumerate(generator):
|
258 |
-
audio_segments.append(audio)
|
259 |
-
|
260 |
-
# Join audio segments using pydub
|
261 |
-
combined = AudioSegment.empty()
|
262 |
-
for seg in audio_segments:
|
263 |
-
segment = AudioSegment(
|
264 |
-
seg.tobytes(),
|
265 |
-
frame_rate=24000,
|
266 |
-
sample_width=seg.dtype.itemsize,
|
267 |
-
channels=1
|
268 |
-
)
|
269 |
-
combined += segment
|
270 |
-
|
271 |
-
output_path = os.path.join(tempfile.gettempdir(), "voice_over_kokoro.wav")
|
272 |
-
combined.export(output_path, format="wav")
|
273 |
-
logging.info(f"Kokoro voice-over generated at {output_path}")
|
274 |
-
return output_path
|
275 |
-
|
276 |
-
except Exception as e:
|
277 |
-
logging.error(f"Error in generate_voice_kokoro: {e}")
|
278 |
-
return f"Error generating Kokoro voice: {e}"
|
279 |
|
280 |
# ---------------------------------------------------------------------
|
281 |
# Music Generation Function
|
282 |
# ---------------------------------------------------------------------
|
283 |
@spaces.GPU(duration=200)
|
284 |
-
def generate_music(prompt: str, audio_length: int)
|
285 |
"""
|
286 |
-
|
287 |
-
|
288 |
-
Args:
|
289 |
-
prompt (str): Music prompt or style suggestion.
|
290 |
-
audio_length (int): Length parameter (number of tokens).
|
291 |
-
|
292 |
-
Returns:
|
293 |
-
str: File path to the generated .wav file or an error message.
|
294 |
"""
|
295 |
try:
|
296 |
if not prompt.strip():
|
297 |
-
|
|
|
298 |
model_key = "facebook/musicgen-large"
|
299 |
musicgen_model, musicgen_processor = get_musicgen_model(model_key)
|
|
|
300 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
301 |
inputs = musicgen_processor(text=[prompt], padding=True, return_tensors="pt").to(device)
|
302 |
|
@@ -305,47 +212,48 @@ def generate_music(prompt: str, audio_length: int) -> Union[str, None]:
|
|
305 |
|
306 |
audio_data = outputs[0, 0].cpu().numpy()
|
307 |
normalized_audio = (audio_data / max(abs(audio_data)) * 32767).astype("int16")
|
308 |
-
|
|
|
309 |
write(output_path, 44100, normalized_audio)
|
310 |
-
|
311 |
return output_path
|
312 |
|
313 |
except Exception as e:
|
314 |
-
logging.error(f"Error in generate_music: {e}")
|
315 |
return f"Error generating music: {e}"
|
316 |
|
|
|
317 |
# ---------------------------------------------------------------------
|
318 |
-
# Audio Blending
|
319 |
# ---------------------------------------------------------------------
|
320 |
@spaces.GPU(duration=100)
|
321 |
-
def blend_audio(voice_path: str, music_path: str, ducking: bool, duck_level: int = 10)
|
322 |
"""
|
323 |
-
|
324 |
-
|
325 |
-
|
326 |
-
|
327 |
-
|
328 |
-
ducking (bool): If True, attenuate music during voice segments.
|
329 |
-
duck_level (int): Attenuation level in dB.
|
330 |
-
|
331 |
-
Returns:
|
332 |
-
str: File path to the blended .wav file or an error message.
|
333 |
"""
|
334 |
try:
|
335 |
-
if not
|
336 |
-
|
337 |
|
338 |
voice = AudioSegment.from_wav(voice_path)
|
339 |
music = AudioSegment.from_wav(music_path)
|
340 |
-
voice_duration = len(voice)
|
341 |
|
342 |
-
|
|
|
|
|
|
|
|
|
343 |
looped_music = AudioSegment.empty()
|
344 |
-
while len(looped_music) <
|
345 |
looped_music += music
|
346 |
music = looped_music
|
347 |
-
|
348 |
-
|
|
|
|
|
349 |
|
350 |
if ducking:
|
351 |
ducked_music = music - duck_level
|
@@ -355,13 +263,12 @@ def blend_audio(voice_path: str, music_path: str, ducking: bool, duck_level: int
|
|
355 |
|
356 |
output_path = os.path.join(tempfile.gettempdir(), "blended_output.wav")
|
357 |
final_audio.export(output_path, format="wav")
|
358 |
-
logging.info(f"Audio blended at {output_path}")
|
359 |
return output_path
|
360 |
|
361 |
except Exception as e:
|
362 |
-
logging.error(f"Error in blend_audio: {e}")
|
363 |
return f"Error blending audio: {e}"
|
364 |
|
|
|
365 |
# ---------------------------------------------------------------------
|
366 |
# Gradio Interface with Enhanced UI
|
367 |
# ---------------------------------------------------------------------
|
@@ -415,7 +322,7 @@ with gr.Blocks(css="""
|
|
415 |
Welcome to **AI Promo Studio**! This platform leverages state-of-the-art AI models to help you generate:
|
416 |
|
417 |
- **Script**: Generate a compelling voice-over script with LLaMA.
|
418 |
-
- **Voice Synthesis**: Create natural-sounding voice-overs using Coqui TTS
|
419 |
- **Music Production**: Produce custom music tracks with MusicGen.
|
420 |
- **Audio Blending**: Seamlessly blend voice and music with options for ducking.
|
421 |
""")
|
@@ -448,26 +355,20 @@ with gr.Blocks(css="""
|
|
448 |
music_suggestion_output = gr.Textbox(label="Music Suggestions", lines=3, interactive=False)
|
449 |
|
450 |
generate_script_button.click(
|
451 |
-
fn=lambda
|
452 |
inputs=[user_prompt, llama_model_id, duration],
|
453 |
outputs=[script_output, sound_design_output, music_suggestion_output],
|
454 |
)
|
455 |
|
456 |
# Step 2: Generate Voice
|
457 |
with gr.Tab("🎤 Voice Synthesis"):
|
458 |
-
gr.Markdown("Generate a natural-sounding voice-over
|
459 |
-
voice_engine = gr.Dropdown(
|
460 |
-
label="TTS Engine",
|
461 |
-
choices=["Coqui TTS", "Kokoro TTS"],
|
462 |
-
value="Coqui TTS",
|
463 |
-
multiselect=False
|
464 |
-
)
|
465 |
selected_tts_model = gr.Dropdown(
|
466 |
-
label="TTS Model
|
467 |
choices=[
|
468 |
-
"tts_models/en/ljspeech/tacotron2-DDC",
|
469 |
-
"tts_models/en/ljspeech/vits",
|
470 |
-
"
|
471 |
],
|
472 |
value="tts_models/en/ljspeech/tacotron2-DDC",
|
473 |
multiselect=False
|
@@ -475,18 +376,9 @@ with gr.Blocks(css="""
|
|
475 |
generate_voice_button = gr.Button("Generate Voice-Over", variant="primary")
|
476 |
voice_audio_output = gr.Audio(label="Voice-Over (WAV)", type="filepath")
|
477 |
|
478 |
-
def generate_voice_combined(script, engine, model_choice):
|
479 |
-
if engine == "Coqui TTS":
|
480 |
-
return generate_voice(script, model_choice)
|
481 |
-
elif engine == "Kokoro TTS":
|
482 |
-
# For Kokoro, pass the voice option (e.g., "af_heart") and default language code ('a')
|
483 |
-
return generate_voice_kokoro(script, lang_code='a', voice=model_choice, speed=1.0)
|
484 |
-
else:
|
485 |
-
return "Error: Unknown TTS engine."
|
486 |
-
|
487 |
generate_voice_button.click(
|
488 |
-
fn=
|
489 |
-
inputs=[script_output,
|
490 |
outputs=voice_audio_output,
|
491 |
)
|
492 |
|
@@ -505,7 +397,7 @@ with gr.Blocks(css="""
|
|
505 |
music_output = gr.Audio(label="Generated Music (WAV)", type="filepath")
|
506 |
|
507 |
generate_music_button.click(
|
508 |
-
fn=lambda
|
509 |
inputs=[music_suggestion_output, audio_length],
|
510 |
outputs=[music_output],
|
511 |
)
|
|
|
1 |
import os
|
2 |
import re
|
|
|
3 |
import torch
|
4 |
import tempfile
|
|
|
5 |
from scipy.io.wavfile import write
|
6 |
from pydub import AudioSegment
|
7 |
from dotenv import load_dotenv
|
|
|
19 |
# Coqui TTS
|
20 |
from TTS.api import TTS
|
21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
# ---------------------------------------------------------------------
|
23 |
+
# Load Environment Variables
|
24 |
# ---------------------------------------------------------------------
|
25 |
load_dotenv()
|
26 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
|
|
|
|
|
|
|
|
27 |
|
28 |
+
# ---------------------------------------------------------------------
|
29 |
# Global Model Caches
|
30 |
+
# ---------------------------------------------------------------------
|
31 |
LLAMA_PIPELINES = {}
|
32 |
MUSICGEN_MODELS = {}
|
33 |
TTS_MODELS = {}
|
34 |
|
35 |
# ---------------------------------------------------------------------
|
36 |
+
# Utility Function: Clean Text
|
37 |
# ---------------------------------------------------------------------
|
38 |
def clean_text(text: str) -> str:
|
39 |
"""
|
40 |
+
Removes undesired characters (e.g., asterisks) that might not be recognized by the model's vocabulary.
|
|
|
|
|
|
|
|
|
|
|
|
|
41 |
"""
|
42 |
+
# Remove all asterisks. You can add more cleaning steps here as needed.
|
43 |
return re.sub(r'\*', '', text)
|
44 |
|
45 |
# ---------------------------------------------------------------------
|
46 |
+
# Helper Functions
|
47 |
# ---------------------------------------------------------------------
|
48 |
+
def get_llama_pipeline(model_id: str, token: str):
|
49 |
"""
|
50 |
+
Returns a cached LLaMA pipeline if available; otherwise, loads it.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
51 |
"""
|
52 |
if model_id in LLAMA_PIPELINES:
|
53 |
return LLAMA_PIPELINES[model_id]
|
54 |
|
55 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id, use_auth_token=token)
|
56 |
+
model = AutoModelForCausalLM.from_pretrained(
|
57 |
+
model_id,
|
58 |
+
use_auth_token=token,
|
59 |
+
torch_dtype=torch.float16,
|
60 |
+
device_map="auto",
|
61 |
+
trust_remote_code=True,
|
62 |
+
)
|
63 |
+
text_pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
64 |
+
LLAMA_PIPELINES[model_id] = text_pipeline
|
65 |
+
return text_pipeline
|
|
|
|
|
|
|
|
|
66 |
|
67 |
+
|
68 |
+
def get_musicgen_model(model_key: str = "facebook/musicgen-large"):
|
69 |
"""
|
70 |
+
Returns a cached MusicGen model if available; otherwise, loads it.
|
71 |
+
Uses the 'large' variant for higher quality outputs.
|
|
|
|
|
|
|
|
|
|
|
72 |
"""
|
73 |
if model_key in MUSICGEN_MODELS:
|
74 |
return MUSICGEN_MODELS[model_key]
|
75 |
|
76 |
+
model = MusicgenForConditionalGeneration.from_pretrained(model_key)
|
77 |
+
processor = AutoProcessor.from_pretrained(model_key)
|
78 |
+
|
79 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
80 |
+
model.to(device)
|
81 |
+
MUSICGEN_MODELS[model_key] = (model, processor)
|
82 |
+
return model, processor
|
|
|
|
|
|
|
83 |
|
84 |
+
|
85 |
+
def get_tts_model(model_name: str = "tts_models/en/ljspeech/tacotron2-DDC"):
|
86 |
"""
|
87 |
+
Returns a cached TTS model if available; otherwise, loads it.
|
|
|
|
|
|
|
|
|
|
|
|
|
88 |
"""
|
89 |
if model_name in TTS_MODELS:
|
90 |
return TTS_MODELS[model_name]
|
91 |
|
92 |
+
tts_model = TTS(model_name)
|
93 |
+
TTS_MODELS[model_name] = tts_model
|
94 |
+
return tts_model
|
95 |
+
|
|
|
|
|
|
|
96 |
|
97 |
# ---------------------------------------------------------------------
|
98 |
# Script Generation Function
|
99 |
# ---------------------------------------------------------------------
|
100 |
@spaces.GPU(duration=100)
|
101 |
+
def generate_script(user_prompt: str, model_id: str, token: str, duration: int):
|
102 |
"""
|
103 |
+
Generates a script, sound design suggestions, and music ideas from a user prompt.
|
104 |
+
Returns a tuple of strings: (voice_script, sound_design, music_suggestions).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
105 |
"""
|
106 |
try:
|
107 |
text_pipeline = get_llama_pipeline(model_id, token)
|
108 |
+
|
109 |
system_prompt = (
|
110 |
"You are an expert radio imaging producer specializing in sound design and music. "
|
111 |
+
f"Based on the user's concept and the selected duration of {duration} seconds, produce the following: "
|
112 |
+
"1. A concise voice-over script. Prefix this section with 'Voice-Over Script:'.\n"
|
113 |
+
"2. Suggestions for sound design. Prefix this section with 'Sound Design Suggestions:'.\n"
|
114 |
+
"3. Music styles or track recommendations. Prefix this section with 'Music Suggestions:'."
|
115 |
)
|
116 |
combined_prompt = f"{system_prompt}\nUser concept: {user_prompt}\nOutput:"
|
117 |
+
|
118 |
with torch.inference_mode():
|
119 |
result = text_pipeline(
|
120 |
combined_prompt,
|
|
|
124 |
)
|
125 |
|
126 |
generated_text = result[0]["generated_text"]
|
|
|
127 |
if "Output:" in generated_text:
|
128 |
generated_text = generated_text.split("Output:")[-1].strip()
|
129 |
|
130 |
+
# Default placeholders
|
131 |
voice_script = "No voice-over script found."
|
132 |
sound_design = "No sound design suggestions found."
|
133 |
music_suggestions = "No music suggestions found."
|
134 |
|
135 |
+
# Voice-Over Script
|
136 |
if "Voice-Over Script:" in generated_text:
|
137 |
+
parts = generated_text.split("Voice-Over Script:")
|
138 |
+
voice_script_part = parts[1]
|
139 |
+
if "Sound Design Suggestions:" in voice_script_part:
|
140 |
+
voice_script = voice_script_part.split("Sound Design Suggestions:")[0].strip()
|
141 |
else:
|
142 |
+
voice_script = voice_script_part.strip()
|
143 |
|
144 |
+
# Sound Design
|
145 |
if "Sound Design Suggestions:" in generated_text:
|
146 |
+
parts = generated_text.split("Sound Design Suggestions:")
|
147 |
+
sound_design_part = parts[1]
|
148 |
+
if "Music Suggestions:" in sound_design_part:
|
149 |
+
sound_design = sound_design_part.split("Music Suggestions:")[0].strip()
|
150 |
else:
|
151 |
+
sound_design = sound_design_part.strip()
|
152 |
|
153 |
+
# Music Suggestions
|
154 |
if "Music Suggestions:" in generated_text:
|
155 |
+
parts = generated_text.split("Music Suggestions:")
|
156 |
+
music_suggestions = parts[1].strip()
|
157 |
|
158 |
return voice_script, sound_design, music_suggestions
|
159 |
|
160 |
except Exception as e:
|
|
|
161 |
return f"Error generating script: {e}", "", ""
|
162 |
|
163 |
+
|
164 |
# ---------------------------------------------------------------------
|
165 |
+
# Voice-Over Generation Function
|
166 |
# ---------------------------------------------------------------------
|
167 |
@spaces.GPU(duration=100)
|
168 |
+
def generate_voice(script: str, tts_model_name: str = "tts_models/en/ljspeech/tacotron2-DDC"):
|
169 |
"""
|
170 |
+
Generates a voice-over from the provided script using the Coqui TTS model.
|
171 |
+
Returns the file path to the generated .wav file.
|
|
|
|
|
|
|
|
|
|
|
|
|
172 |
"""
|
173 |
try:
|
174 |
if not script.strip():
|
175 |
+
return "Error: No script provided."
|
176 |
+
|
177 |
+
# Clean the script to remove special characters (e.g., asterisks) that may produce warnings
|
178 |
cleaned_script = clean_text(script)
|
179 |
+
|
180 |
tts_model = get_tts_model(tts_model_name)
|
181 |
+
|
182 |
+
# Generate and save voice
|
183 |
+
output_path = os.path.join(tempfile.gettempdir(), "voice_over.wav")
|
184 |
tts_model.tts_to_file(text=cleaned_script, file_path=output_path)
|
|
|
185 |
return output_path
|
186 |
|
187 |
except Exception as e:
|
|
|
188 |
return f"Error generating voice: {e}"
|
189 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
190 |
|
191 |
# ---------------------------------------------------------------------
|
192 |
# Music Generation Function
|
193 |
# ---------------------------------------------------------------------
|
194 |
@spaces.GPU(duration=200)
|
195 |
+
def generate_music(prompt: str, audio_length: int):
|
196 |
"""
|
197 |
+
Generates music from the 'facebook/musicgen-large' model based on the prompt.
|
198 |
+
Returns the file path to the generated .wav file.
|
|
|
|
|
|
|
|
|
|
|
|
|
199 |
"""
|
200 |
try:
|
201 |
if not prompt.strip():
|
202 |
+
return "Error: No music suggestion provided."
|
203 |
+
|
204 |
model_key = "facebook/musicgen-large"
|
205 |
musicgen_model, musicgen_processor = get_musicgen_model(model_key)
|
206 |
+
|
207 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
208 |
inputs = musicgen_processor(text=[prompt], padding=True, return_tensors="pt").to(device)
|
209 |
|
|
|
212 |
|
213 |
audio_data = outputs[0, 0].cpu().numpy()
|
214 |
normalized_audio = (audio_data / max(abs(audio_data)) * 32767).astype("int16")
|
215 |
+
|
216 |
+
output_path = os.path.join(tempfile.gettempdir(), "musicgen_large_generated_music.wav")
|
217 |
write(output_path, 44100, normalized_audio)
|
218 |
+
|
219 |
return output_path
|
220 |
|
221 |
except Exception as e:
|
|
|
222 |
return f"Error generating music: {e}"
|
223 |
|
224 |
+
|
225 |
# ---------------------------------------------------------------------
|
226 |
+
# Audio Blending with Duration Sync & Ducking
|
227 |
# ---------------------------------------------------------------------
|
228 |
@spaces.GPU(duration=100)
|
229 |
+
def blend_audio(voice_path: str, music_path: str, ducking: bool, duck_level: int = 10):
|
230 |
"""
|
231 |
+
Blends two audio files (voice and music).
|
232 |
+
1. If music < voice, loops the music until it meets/exceeds the voice duration.
|
233 |
+
2. If music > voice, trims music to the voice duration.
|
234 |
+
3. If ducking=True, the music is attenuated by 'duck_level' dB while the voice is playing.
|
235 |
+
Returns the file path to the blended .wav file.
|
|
|
|
|
|
|
|
|
|
|
236 |
"""
|
237 |
try:
|
238 |
+
if not os.path.isfile(voice_path) or not os.path.isfile(music_path):
|
239 |
+
return "Error: Missing audio files for blending."
|
240 |
|
241 |
voice = AudioSegment.from_wav(voice_path)
|
242 |
music = AudioSegment.from_wav(music_path)
|
|
|
243 |
|
244 |
+
voice_len = len(voice) # in milliseconds
|
245 |
+
music_len = len(music) # in milliseconds
|
246 |
+
|
247 |
+
# Loop music if it's shorter than the voice
|
248 |
+
if music_len < voice_len:
|
249 |
looped_music = AudioSegment.empty()
|
250 |
+
while len(looped_music) < voice_len:
|
251 |
looped_music += music
|
252 |
music = looped_music
|
253 |
+
|
254 |
+
# Trim music if it's longer than the voice
|
255 |
+
if len(music) > voice_len:
|
256 |
+
music = music[:voice_len]
|
257 |
|
258 |
if ducking:
|
259 |
ducked_music = music - duck_level
|
|
|
263 |
|
264 |
output_path = os.path.join(tempfile.gettempdir(), "blended_output.wav")
|
265 |
final_audio.export(output_path, format="wav")
|
|
|
266 |
return output_path
|
267 |
|
268 |
except Exception as e:
|
|
|
269 |
return f"Error blending audio: {e}"
|
270 |
|
271 |
+
|
272 |
# ---------------------------------------------------------------------
|
273 |
# Gradio Interface with Enhanced UI
|
274 |
# ---------------------------------------------------------------------
|
|
|
322 |
Welcome to **AI Promo Studio**! This platform leverages state-of-the-art AI models to help you generate:
|
323 |
|
324 |
- **Script**: Generate a compelling voice-over script with LLaMA.
|
325 |
+
- **Voice Synthesis**: Create natural-sounding voice-overs using Coqui TTS.
|
326 |
- **Music Production**: Produce custom music tracks with MusicGen.
|
327 |
- **Audio Blending**: Seamlessly blend voice and music with options for ducking.
|
328 |
""")
|
|
|
355 |
music_suggestion_output = gr.Textbox(label="Music Suggestions", lines=3, interactive=False)
|
356 |
|
357 |
generate_script_button.click(
|
358 |
+
fn=lambda user_prompt, model_id, dur: generate_script(user_prompt, model_id, HF_TOKEN, dur),
|
359 |
inputs=[user_prompt, llama_model_id, duration],
|
360 |
outputs=[script_output, sound_design_output, music_suggestion_output],
|
361 |
)
|
362 |
|
363 |
# Step 2: Generate Voice
|
364 |
with gr.Tab("🎤 Voice Synthesis"):
|
365 |
+
gr.Markdown("Generate a natural-sounding voice-over using Coqui TTS.")
|
|
|
|
|
|
|
|
|
|
|
|
|
366 |
selected_tts_model = gr.Dropdown(
|
367 |
+
label="TTS Model",
|
368 |
choices=[
|
369 |
+
"tts_models/en/ljspeech/tacotron2-DDC",
|
370 |
+
"tts_models/en/ljspeech/vits",
|
371 |
+
"tts_models/en/sam/tacotron-DDC",
|
372 |
],
|
373 |
value="tts_models/en/ljspeech/tacotron2-DDC",
|
374 |
multiselect=False
|
|
|
376 |
generate_voice_button = gr.Button("Generate Voice-Over", variant="primary")
|
377 |
voice_audio_output = gr.Audio(label="Voice-Over (WAV)", type="filepath")
|
378 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
379 |
generate_voice_button.click(
|
380 |
+
fn=lambda script, tts_model: generate_voice(script, tts_model),
|
381 |
+
inputs=[script_output, selected_tts_model],
|
382 |
outputs=voice_audio_output,
|
383 |
)
|
384 |
|
|
|
397 |
music_output = gr.Audio(label="Generated Music (WAV)", type="filepath")
|
398 |
|
399 |
generate_music_button.click(
|
400 |
+
fn=lambda music_suggestion, length: generate_music(music_suggestion, length),
|
401 |
inputs=[music_suggestion_output, audio_length],
|
402 |
outputs=[music_output],
|
403 |
)
|