Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -45,10 +45,10 @@ class ModelManager:
|
|
45 |
|
46 |
def get_llama_pipeline(self, model_id, token):
|
47 |
if model_id not in self.llama_pipelines:
|
48 |
-
tokenizer = AutoTokenizer.from_pretrained(model_id,
|
49 |
model = AutoModelForCausalLM.from_pretrained(
|
50 |
model_id,
|
51 |
-
|
52 |
torch_dtype=torch.float16,
|
53 |
device_map="auto",
|
54 |
attn_implementation="flash_attention_2"
|
@@ -85,20 +85,17 @@ def generate_script(user_prompt, model_id, duration, temperature=0.7, max_tokens
|
|
85 |
try:
|
86 |
text_pipeline = model_manager.get_llama_pipeline(model_id, HF_TOKEN)
|
87 |
|
88 |
-
system_prompt = f"""You are an
|
89 |
-
1. Voice Script: [Clear
|
90 |
-
2. Sound Design: [3-5
|
91 |
-
3. Music: [Genre
|
92 |
|
93 |
-
|
94 |
-
|
95 |
-
messages = [
|
96 |
-
{"role": "system", "content": system_prompt},
|
97 |
-
{"role": "user", "content": user_prompt}
|
98 |
-
]
|
99 |
|
|
|
|
|
100 |
response = text_pipeline(
|
101 |
-
|
102 |
max_new_tokens=max_tokens,
|
103 |
temperature=temperature,
|
104 |
do_sample=True,
|
@@ -106,8 +103,7 @@ Keep sections concise and production-ready."""
|
|
106 |
eos_token_id=text_pipeline.tokenizer.eos_token_id
|
107 |
)
|
108 |
|
109 |
-
return parse_generated_content(response[0][
|
110 |
-
|
111 |
except Exception as e:
|
112 |
return f"Error: {str(e)}", "", ""
|
113 |
|
@@ -140,10 +136,10 @@ def parse_generated_content(text):
|
|
140 |
def generate_voice(script, tts_model, speed=1.0):
|
141 |
try:
|
142 |
if not script.strip():
|
143 |
-
|
144 |
|
145 |
tts = model_manager.get_tts_model(tts_model)
|
146 |
-
output_path = os.path.join(tempfile.gettempdir(), "
|
147 |
|
148 |
tts.tts_to_file(
|
149 |
text=script,
|
@@ -174,8 +170,8 @@ def generate_music(prompt, duration_sec=30, temperature=1.0, guidance_scale=3.0)
|
|
174 |
do_sample=True
|
175 |
)
|
176 |
|
177 |
-
output_path = os.path.join(tempfile.gettempdir(), "
|
178 |
-
write(output_path,
|
179 |
return output_path
|
180 |
except Exception as e:
|
181 |
return f"Error: {str(e)}"
|
@@ -185,19 +181,22 @@ def blend_audio(voice_path, music_path, ducking=True, duck_level=10, crossfade=5
|
|
185 |
voice = AudioSegment.from_wav(voice_path)
|
186 |
music = AudioSegment.from_wav(music_path)
|
187 |
|
|
|
188 |
if len(music) < len(voice):
|
189 |
loops = (len(voice) // len(music)) + 1
|
190 |
music = music * loops
|
191 |
|
192 |
music = music[:len(voice)].fade_out(crossfade)
|
|
|
193 |
|
|
|
194 |
if ducking:
|
195 |
ducked_music = music - duck_level
|
196 |
-
mixed = ducked_music.overlay(voice
|
197 |
else:
|
198 |
mixed = music.overlay(voice)
|
199 |
|
200 |
-
output_path = os.path.join(tempfile.gettempdir(), "
|
201 |
mixed.export(output_path, format="wav")
|
202 |
return output_path
|
203 |
except Exception as e:
|
@@ -214,124 +213,119 @@ theme = gr.themes.Soft(
|
|
214 |
background_fill_primary_dark='#1F1F1F'
|
215 |
)
|
216 |
|
217 |
-
with gr.Blocks(theme=theme, title="AI
|
218 |
gr.Markdown("""
|
219 |
-
#
|
220 |
-
*
|
221 |
""")
|
222 |
-
|
223 |
with gr.Tabs():
|
224 |
-
|
|
|
225 |
with gr.Row():
|
226 |
with gr.Column(scale=2):
|
227 |
concept_input = gr.Textbox(
|
228 |
-
label="Your
|
229 |
-
placeholder="
|
230 |
-
lines=3
|
231 |
-
max_lines=6
|
232 |
)
|
233 |
with gr.Accordion("Advanced Settings", open=False):
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
|
241 |
-
|
242 |
-
temp_slider = gr.Slider(0.1, 1.5, value=0.7, step=0.1, label="Creativity")
|
243 |
-
token_slider = gr.Slider(128, 1024, value=512, step=128, label="Max Length")
|
244 |
-
|
245 |
-
generate_btn = gr.Button("✨ Generate Concept", variant="primary")
|
246 |
|
247 |
with gr.Column(scale=1):
|
248 |
script_output = gr.Textbox(label="Voice Script", interactive=True)
|
249 |
sound_output = gr.Textbox(label="Sound Design", interactive=True)
|
250 |
-
music_output = gr.Textbox(label="Music
|
251 |
-
|
252 |
-
generate_btn.click(
|
253 |
-
generate_script,
|
254 |
-
inputs=[concept_input, model_selector, duration_slider, temp_slider, token_slider],
|
255 |
-
outputs=[script_output, sound_output, music_output]
|
256 |
-
)
|
257 |
|
258 |
-
|
|
|
259 |
with gr.Row():
|
260 |
with gr.Column():
|
261 |
-
|
262 |
choices=list(MODEL_CONFIG["tts_models"].values()),
|
263 |
label="Voice Model",
|
264 |
-
value=
|
265 |
)
|
266 |
-
speed_slider = gr.Slider(0.5, 2.0,
|
267 |
-
voice_btn = gr.Button("
|
268 |
with gr.Column():
|
269 |
-
voice_preview = gr.Audio(label="Preview",
|
270 |
-
voice_btn.click(
|
271 |
-
generate_voice,
|
272 |
-
inputs=[script_output, tts_model, speed_slider],
|
273 |
-
outputs=voice_preview
|
274 |
-
)
|
275 |
|
276 |
-
|
|
|
277 |
with gr.Row():
|
278 |
with gr.Column():
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
music_btn = gr.Button("🎵 Generate Music", variant="primary")
|
284 |
with gr.Column():
|
285 |
-
music_preview = gr.Audio(label="Preview",
|
286 |
-
music_btn.click(
|
287 |
-
generate_music,
|
288 |
-
inputs=[music_output, music_duration, music_temp, guidance_scale],
|
289 |
-
outputs=music_preview
|
290 |
-
)
|
291 |
|
292 |
-
|
|
|
293 |
with gr.Row():
|
294 |
with gr.Column():
|
295 |
-
ducking_toggle = gr.Checkbox(
|
296 |
-
duck_level = gr.Slider(0, 30,
|
297 |
-
crossfade_time = gr.Slider(0, 2000,
|
298 |
-
mix_btn = gr.Button("
|
299 |
with gr.Column():
|
300 |
-
final_mix = gr.Audio(label="Master Output",
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
outputs=final_mix
|
305 |
-
)
|
306 |
-
|
307 |
-
with gr.Accordion("📚 Example Prompts", open=False):
|
308 |
gr.Examples(
|
309 |
examples=[
|
310 |
-
["A
|
311 |
-
["A 15-second radio ad
|
312 |
["A 60-second documentary trailer with epic orchestral music"]
|
313 |
],
|
314 |
inputs=concept_input
|
315 |
)
|
316 |
-
|
317 |
-
with gr.Row():
|
318 |
-
gr.Markdown("### System Resources")
|
319 |
-
gpu_status = gr.Textbox(label="GPU Utilization", interactive=False)
|
320 |
-
ram_status = gr.Textbox(label="RAM Usage", interactive=False)
|
321 |
|
322 |
-
# Custom Footer
|
323 |
gr.Markdown("""
|
324 |
-
<
|
325 |
-
|
326 |
-
|
327 |
-
|
|
|
|
|
|
|
|
|
328 |
""")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
329 |
|
330 |
-
|
331 |
-
|
332 |
-
|
333 |
-
|
334 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
335 |
|
336 |
if __name__ == "__main__":
|
337 |
demo.launch(server_name="0.0.0.0", server_port=7860)
|
|
|
45 |
|
46 |
def get_llama_pipeline(self, model_id, token):
|
47 |
if model_id not in self.llama_pipelines:
|
48 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id, use_auth_token=token)
|
49 |
model = AutoModelForCausalLM.from_pretrained(
|
50 |
model_id,
|
51 |
+
use_auth_token=token,
|
52 |
torch_dtype=torch.float16,
|
53 |
device_map="auto",
|
54 |
attn_implementation="flash_attention_2"
|
|
|
85 |
try:
|
86 |
text_pipeline = model_manager.get_llama_pipeline(model_id, HF_TOKEN)
|
87 |
|
88 |
+
system_prompt = f"""You are an expert radio imaging producer. Create content for a {duration}-second promo:
|
89 |
+
1. Voice Script: [Clear narration]
|
90 |
+
2. Sound Design: [3-5 effects]
|
91 |
+
3. Music: [Genre/tempo/mood]
|
92 |
|
93 |
+
Respond in this exact format:"""
|
|
|
|
|
|
|
|
|
|
|
94 |
|
95 |
+
prompt = f"{system_prompt}\nConcept: {user_prompt}\nVoice Script:"
|
96 |
+
|
97 |
response = text_pipeline(
|
98 |
+
prompt,
|
99 |
max_new_tokens=max_tokens,
|
100 |
temperature=temperature,
|
101 |
do_sample=True,
|
|
|
103 |
eos_token_id=text_pipeline.tokenizer.eos_token_id
|
104 |
)
|
105 |
|
106 |
+
return parse_generated_content(response[0]["generated_text"])
|
|
|
107 |
except Exception as e:
|
108 |
return f"Error: {str(e)}", "", ""
|
109 |
|
|
|
136 |
def generate_voice(script, tts_model, speed=1.0):
|
137 |
try:
|
138 |
if not script.strip():
|
139 |
+
return "Error: Empty script"
|
140 |
|
141 |
tts = model_manager.get_tts_model(tts_model)
|
142 |
+
output_path = os.path.join(tempfile.gettempdir(), "voice.wav")
|
143 |
|
144 |
tts.tts_to_file(
|
145 |
text=script,
|
|
|
170 |
do_sample=True
|
171 |
)
|
172 |
|
173 |
+
output_path = os.path.join(tempfile.gettempdir(), "music.wav")
|
174 |
+
write(output_path, 44100, audio_values[0, 0].cpu().numpy())
|
175 |
return output_path
|
176 |
except Exception as e:
|
177 |
return f"Error: {str(e)}"
|
|
|
181 |
voice = AudioSegment.from_wav(voice_path)
|
182 |
music = AudioSegment.from_wav(music_path)
|
183 |
|
184 |
+
# Align durations with crossfade
|
185 |
if len(music) < len(voice):
|
186 |
loops = (len(voice) // len(music)) + 1
|
187 |
music = music * loops
|
188 |
|
189 |
music = music[:len(voice)].fade_out(crossfade)
|
190 |
+
voice = voice.fade_in(crossfade)
|
191 |
|
192 |
+
# Apply ducking
|
193 |
if ducking:
|
194 |
ducked_music = music - duck_level
|
195 |
+
mixed = ducked_music.overlay(voice)
|
196 |
else:
|
197 |
mixed = music.overlay(voice)
|
198 |
|
199 |
+
output_path = os.path.join(tempfile.gettempdir(), "final_mix.wav")
|
200 |
mixed.export(output_path, format="wav")
|
201 |
return output_path
|
202 |
except Exception as e:
|
|
|
213 |
background_fill_primary_dark='#1F1F1F'
|
214 |
)
|
215 |
|
216 |
+
with gr.Blocks(theme=theme, title="AI Radio Studio Pro") as demo:
|
217 |
gr.Markdown("""
|
218 |
+
# 🎧 AI Radio Studio Pro
|
219 |
+
*Professional Audio Production in 4 Steps*
|
220 |
""")
|
221 |
+
|
222 |
with gr.Tabs():
|
223 |
+
# Step 1: Concept Development
|
224 |
+
with gr.Tab("1️⃣ Concept"):
|
225 |
with gr.Row():
|
226 |
with gr.Column(scale=2):
|
227 |
concept_input = gr.Textbox(
|
228 |
+
label="Your Idea",
|
229 |
+
placeholder="e.g., A 30-second morning show intro with energetic music...",
|
230 |
+
lines=3
|
|
|
231 |
)
|
232 |
with gr.Accordion("Advanced Settings", open=False):
|
233 |
+
model_selector = gr.Dropdown(
|
234 |
+
choices=list(MODEL_CONFIG["llama_models"].values()),
|
235 |
+
label="AI Model",
|
236 |
+
value=MODEL_CONFIG["llama_models"]["Meta-Llama-3-8B"]
|
237 |
+
)
|
238 |
+
duration_slider = gr.Slider(15, 120, 30, step=15, label="Duration (seconds)")
|
239 |
+
temp_slider = gr.Slider(0.1, 1.5, 0.7, step=0.1, label="Creativity")
|
240 |
+
generate_btn = gr.Button("Generate Script", variant="primary")
|
|
|
|
|
|
|
|
|
241 |
|
242 |
with gr.Column(scale=1):
|
243 |
script_output = gr.Textbox(label="Voice Script", interactive=True)
|
244 |
sound_output = gr.Textbox(label="Sound Design", interactive=True)
|
245 |
+
music_output = gr.Textbox(label="Music Style", interactive=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
246 |
|
247 |
+
# Step 2: Voice Production
|
248 |
+
with gr.Tab("2️⃣ Voice"):
|
249 |
with gr.Row():
|
250 |
with gr.Column():
|
251 |
+
tts_selector = gr.Dropdown(
|
252 |
choices=list(MODEL_CONFIG["tts_models"].values()),
|
253 |
label="Voice Model",
|
254 |
+
value="tts_models/en/ljspeech/tacotron2-DDC"
|
255 |
)
|
256 |
+
speed_slider = gr.Slider(0.5, 2.0, 1.0, step=0.1, label="Speaking Rate")
|
257 |
+
voice_btn = gr.Button("Generate Voiceover", variant="primary")
|
258 |
with gr.Column():
|
259 |
+
voice_preview = gr.Audio(label="Preview", type="filepath")
|
|
|
|
|
|
|
|
|
|
|
260 |
|
261 |
+
# Step 3: Music Production
|
262 |
+
with gr.Tab("3️⃣ Music"):
|
263 |
with gr.Row():
|
264 |
with gr.Column():
|
265 |
+
music_duration = gr.Slider(10, 120, 30, label="Duration (seconds)")
|
266 |
+
music_temp = gr.Slider(0.1, 2.0, 1.0, label="Creativity")
|
267 |
+
guidance_scale = gr.Slider(1.0, 5.0, 3.0, label="Focus")
|
268 |
+
music_btn = gr.Button("Generate Music", variant="primary")
|
|
|
269 |
with gr.Column():
|
270 |
+
music_preview = gr.Audio(label="Preview", type="filepath")
|
|
|
|
|
|
|
|
|
|
|
271 |
|
272 |
+
# Step 4: Final Mix
|
273 |
+
with gr.Tab("4️⃣ Mix"):
|
274 |
with gr.Row():
|
275 |
with gr.Column():
|
276 |
+
ducking_toggle = gr.Checkbox(True, label="Enable Voice Ducking")
|
277 |
+
duck_level = gr.Slider(0, 30, 12, label="Ducking Strength (dB)")
|
278 |
+
crossfade_time = gr.Slider(0, 2000, 500, label="Crossfade (ms)")
|
279 |
+
mix_btn = gr.Button("Create Final Mix", variant="primary")
|
280 |
with gr.Column():
|
281 |
+
final_mix = gr.Audio(label="Master Output", type="filepath")
|
282 |
+
|
283 |
+
# Examples & Footer
|
284 |
+
with gr.Accordion("💡 Example Prompts", open=False):
|
|
|
|
|
|
|
|
|
285 |
gr.Examples(
|
286 |
examples=[
|
287 |
+
["A 45-second tech podcast intro with futuristic synth effects"],
|
288 |
+
["A 15-second coffee shop radio ad with morning acoustic vibes"],
|
289 |
["A 60-second documentary trailer with epic orchestral music"]
|
290 |
],
|
291 |
inputs=concept_input
|
292 |
)
|
|
|
|
|
|
|
|
|
|
|
293 |
|
|
|
294 |
gr.Markdown("""
|
295 |
+
<div style="text-align: center; margin-top: 30px; padding-top: 20px; border-top: 1px solid #444;">
|
296 |
+
<p style="font-size: 0.9em; color: #888;">
|
297 |
+
Created with ❤️ by <a href="https://bilsimaging.com" target="_blank" style="color: #66b3ff;">bilsimaging.com</a>
|
298 |
+
</p>
|
299 |
+
<a href="https://visitorbadge.io/status?path=https%3A%2F%2Fhuggingface.co%2Fspaces%2FBils%2Fradiogold">
|
300 |
+
<img src="https://api.visitorbadge.io/api/visitors?path=https%3A%2F%2Fhuggingface.co%2Fspaces%2FBils%2Fradiogold&countColor=%23263759"/>
|
301 |
+
</a>
|
302 |
+
</div>
|
303 |
""")
|
304 |
+
|
305 |
+
# Event Handling
|
306 |
+
generate_btn.click(
|
307 |
+
generate_script,
|
308 |
+
inputs=[concept_input, model_selector, duration_slider, temp_slider],
|
309 |
+
outputs=[script_output, sound_output, music_output]
|
310 |
+
)
|
311 |
|
312 |
+
voice_btn.click(
|
313 |
+
generate_voice,
|
314 |
+
inputs=[script_output, tts_selector, speed_slider],
|
315 |
+
outputs=voice_preview
|
316 |
+
)
|
317 |
+
|
318 |
+
music_btn.click(
|
319 |
+
generate_music,
|
320 |
+
inputs=[music_output, music_duration, music_temp, guidance_scale],
|
321 |
+
outputs=music_preview
|
322 |
+
)
|
323 |
+
|
324 |
+
mix_btn.click(
|
325 |
+
blend_audio,
|
326 |
+
inputs=[voice_preview, music_preview, ducking_toggle, duck_level, crossfade_time],
|
327 |
+
outputs=final_mix
|
328 |
+
)
|
329 |
|
330 |
if __name__ == "__main__":
|
331 |
demo.launch(server_name="0.0.0.0", server_port=7860)
|