t

.gitignore

@@ -0,0 +1 @@
+__pycache__

app.py

@@ -7,9 +7,11 @@ import soundfile as sf
 import tempfile
 import os
 import re
+import spaces
 
 # Global variables to persist across requests
 device = "cuda" if torch.cuda.is_available() else "cpu"
+print ("Device:", device)
 model = None
 processor = None
 current_model_id = None
@@ -145,6 +147,7 @@ def load_audio_file(filepath, target_sr=24000):
         print(f"Error loading audio {filepath}: {e}")
         return None
 
+@spaces.GPU
 def generate_simple_audio(text, speaker_id):
     """Generate simple text-to-speech"""
     global model, processor
@@ -170,6 +173,7 @@ def generate_simple_audio(text, speaker_id):
     except Exception as e:
         return None, f"Error: {str(e)}"
 
+@spaces.GPU
 def generate_context_audio(text, speaker_id, context_text, context_audio_file):
     """Generate audio with context"""
     global model, processor
@@ -250,6 +254,7 @@ def extract_audio_output(audio_output):
         print(f"Error extracting audio: {e}")
         return np.array([])
 
+@spaces.GPU
 def generate_conversation(speaker_a, speaker_b, speaker_a_audio, speaker_a_text,
                         speaker_b_audio, speaker_b_text, dialogue_text,
                         split_sentences_flag, use_style, progress=gr.Progress()):
@@ -390,15 +395,16 @@ def generate_conversation(speaker_a, speaker_b, speaker_a_audio, speaker_a_text,
         return None, f"Error: {str(e)}"
 
 # Initialize with default model
+@spaces.GPU
 def initialize_default_model():
     """Initialize the default model on startup"""
     default_model_name = list(MODELS.keys())[0]
     load_model(default_model_name)
 
 # Create interface
-def create_interface():
-    with gr.Blocks(
-        title="CSM Dhivehi Speech Generator",
+
+with gr.Blocks(
+        title="Dhivehi TTS Demos",
         css="""
         .dhivehi-text textarea {
             font-size: 18px !important;
@@ -415,237 +421,235 @@ def create_interface():
         }
         """
     ) as app:
-        gr.Markdown("# 🎙️  CSM Dhivehi Speech Generator")
-        
-        # Model selection
-        with gr.Row():
-            model_dropdown = gr.Dropdown(
-                choices=list(MODELS.keys()),
-                value=list(MODELS.keys())[0],
-                label="🤖 Select Model"
-            )
-            model_info = gr.Textbox(
-                value=get_model_info(),
-                label="Model Status",
-                interactive=False
-            )
-        
-        def change_model(model_name):
-            success = load_model(model_name)
-            if success:
-                choices = get_speaker_choices()
-                # Generate new speaker display choices for the new model
-                current_prompts = get_current_speaker_prompts()
-                new_speaker_choices = [(f"{prompt_data['speaker_id']}: {prompt_data['name']}", prompt_data['speaker_id'])
-                                     for prompt_key, prompt_data in current_prompts.items()]
-                display_choices = [choice[0] for choice in new_speaker_choices]
-                
-                return (
-                    gr.update(value=get_model_info()),
-                    gr.update(choices=choices, value=choices[0]),
-                    gr.update(choices=choices, value=choices[1] if len(choices) > 1 else choices[0]),
-                    # Update simple generation speaker radio
-                    gr.update(choices=display_choices, value=display_choices[0] if display_choices else "0: Speaker"),
-                    # Update context generation speaker radio
-                    gr.update(choices=display_choices, value=display_choices[0] if display_choices else "0: Speaker")
+        with gr.Tab("🎙️ CSM Speech Generator"):
+            gr.Markdown("# 🎙️  CSM Dhivehi Speech Generator")
+            
+            # Model selection
+            with gr.Row():
+                model_dropdown = gr.Dropdown(
+                    choices=list(MODELS.keys()),
+                    value=list(MODELS.keys())[0],
+                    label="🤖 Select Model"
                 )
-            return (
-                gr.update(value="Error loading model"),
-                gr.update(),
-                gr.update(),
-                gr.update(),
-                gr.update()
-            )
-        
-        with gr.Tabs():
-            # Simple Generation
-            with gr.TabItem("🎯 Simple Generation"):
-                gr.Markdown("### Generate speech from text without context")
-                
-                with gr.Row():
-                    with gr.Column():
-                        simple_text = gr.Textbox(
-                            label="Text to Generate (Dhivehi)",
-                            placeholder="މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ",
-                            value="މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ",
-                            lines=3,
-                            elem_classes=["dhivehi-text"]
-                        )
-                        # Generate speaker choices from current model's speaker prompts
-                        current_prompts = get_current_speaker_prompts()
-                        speaker_choices = [(f"{prompt_data['speaker_id']}: {prompt_data['name']}", prompt_data['speaker_id'])
-                                         for prompt_key, prompt_data in current_prompts.items()]
-                        simple_speaker = gr.Radio(
-                            choices=[choice[0] for choice in speaker_choices],
-                            label="Speaker",
-                            value=speaker_choices[0][0] if speaker_choices else "0: Speaker"
-                        )
-                        simple_btn = gr.Button("🎵 Generate", variant="primary")
-                    
-                    with gr.Column():
-                        simple_audio = gr.Audio(label="Generated Audio")
-                        simple_status = gr.Textbox(label="Status", interactive=False)
-                
-                def simple_generate_with_mapping(text, speaker_display):
-                    # Extract speaker ID from display text (e.g., "0: Female Speaker 01" -> "0")
-                    speaker_id = speaker_display.split(":")[0]
-                    return generate_simple_audio(text, speaker_id)
-                
-                simple_btn.click(
-                    simple_generate_with_mapping,
-                    inputs=[simple_text, simple_speaker],
-                    outputs=[simple_audio, simple_status]
+                model_info = gr.Textbox(
+                    value=get_model_info(),
+                    label="Model Status",
+                    interactive=False
                 )
             
-            # Context Generation
-            with gr.TabItem("🎭 Context Generation"):
-                gr.Markdown("### Generate speech with audio context")
-                
-                with gr.Row():
-                    with gr.Column():
-                        context_text = gr.Textbox(
-                            label="Context Text",
-                            placeholder="މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ",
-                            value="އައިލީޝް އޭނާއަތަށް ލިބުނީ އެންމެ ނުވަ މަހުގަ",
-                            lines=2,
-                            elem_classes=["dhivehi-text"]
-                        )
-                        context_audio = gr.Audio(
-                            label="Context Audio",
-                            type="filepath"
-                        )
-                        target_text = gr.Textbox(
-                            label="Text to Generate",
-                            placeholder="މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ",
-                            value="މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ",
-                            lines=3,
-                            elem_classes=["dhivehi-text"]
-                        )
-                        # Generate speaker choices for context generation
-                        context_speaker = gr.Radio(
-                            choices=[choice[0] for choice in speaker_choices],
-                            label="Speaker",
-                            value=speaker_choices[0][0] if speaker_choices else "0: Speaker"
-                        )
-                        context_btn = gr.Button("🎵 Generate with Context", variant="primary")
+            def change_model(model_name):
+                success = load_model(model_name)
+                if success:
+                    choices = get_speaker_choices()
+                    # Generate new speaker display choices for the new model
+                    current_prompts = get_current_speaker_prompts()
+                    new_speaker_choices = [(f"{prompt_data['speaker_id']}: {prompt_data['name']}", prompt_data['speaker_id'])
+                                        for prompt_key, prompt_data in current_prompts.items()]
+                    display_choices = [choice[0] for choice in new_speaker_choices]
                     
-                    with gr.Column():
-                        context_audio_out = gr.Audio(label="Generated Audio")
-                        context_status = gr.Textbox(label="Status", interactive=False)
-                
-                def context_generate_with_mapping(text, speaker_display, context_text_val, context_audio_val):
-                    # Extract speaker ID from display text
-                    speaker_id = speaker_display.split(":")[0]
-                    return generate_context_audio(text, speaker_id, context_text_val, context_audio_val)
-                
-                context_btn.click(
-                    context_generate_with_mapping,
-                    inputs=[target_text, context_speaker, context_text, context_audio],
-                    outputs=[context_audio_out, context_status]
+                    return (
+                        gr.update(value=get_model_info()),
+                        gr.update(choices=choices, value=choices[0]),
+                        gr.update(choices=choices, value=choices[1] if len(choices) > 1 else choices[0]),
+                        # Update simple generation speaker radio
+                        gr.update(choices=display_choices, value=display_choices[0] if display_choices else "0: Speaker"),
+                        # Update context generation speaker radio
+                        gr.update(choices=display_choices, value=display_choices[0] if display_choices else "0: Speaker")
+                    )
+                return (
+                    gr.update(value="Error loading model"),
+                    gr.update(),
+                    gr.update(),
+                    gr.update(),
+                    gr.update()
                 )
             
-            # Conversation Generation
-            with gr.TabItem("💬 Conversation"):
-                gr.Markdown("### Generate dual-speaker conversations")
-                
-                with gr.Row():
-                    speaker_a = gr.Dropdown(
-                        choices=get_speaker_choices(),
-                        label="Speaker A",
-                        value=get_speaker_choices()[0]
-                    )
-                    speaker_b = gr.Dropdown(
-                        choices=get_speaker_choices(),
-                        label="Speaker B",
-                        value=get_speaker_choices()[1] if len(get_speaker_choices()) > 1 else get_speaker_choices()[0]
-                    )
-                
-                with gr.Accordion("🎵 Audio Style References", open=False):
+            with gr.Tabs():
+                # Simple Generation
+                with gr.TabItem("🎯 Simple Generation"):
+                    gr.Markdown("### Generate speech from text without context")
+                    
                     with gr.Row():
                         with gr.Column():
-                            gr.Markdown("**Speaker A Context**")
-                            speaker_a_audio = gr.Audio(
-                                type="filepath",
-                                label="Speaker A Audio Style"
-                            )
-                            speaker_a_text = gr.Textbox(
-                                label="Speaker A Context Text",
-                                lines=2,
+                            simple_text = gr.Textbox(
+                                label="Text to Generate (Dhivehi)",
                                 placeholder="މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ",
+                                value="މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ",
+                                lines=3,
                                 elem_classes=["dhivehi-text"]
                             )
+                            # Generate speaker choices from current model's speaker prompts
+                            current_prompts = get_current_speaker_prompts()
+                            speaker_choices = [(f"{prompt_data['speaker_id']}: {prompt_data['name']}", prompt_data['speaker_id'])
+                                            for prompt_key, prompt_data in current_prompts.items()]
+                            simple_speaker = gr.Radio(
+                                choices=[choice[0] for choice in speaker_choices],
+                                label="Speaker",
+                                value=speaker_choices[0][0] if speaker_choices else "0: Speaker"
+                            )
+                            simple_btn = gr.Button("🎵 Generate", variant="primary")
                         
                         with gr.Column():
-                            gr.Markdown("**Speaker B Context**")
-                            speaker_b_audio = gr.Audio(
-                                type="filepath",
-                                label="Speaker B Audio Style"
-                            )
-                            speaker_b_text = gr.Textbox(
-                                label="Speaker B Context Text",
+                            simple_audio = gr.Audio(label="Generated Audio")
+                            simple_status = gr.Textbox(label="Status", interactive=False)
+                    
+                    def simple_generate_with_mapping(text, speaker_display):
+                        # Extract speaker ID from display text (e.g., "0: Female Speaker 01" -> "0")
+                        speaker_id = speaker_display.split(":")[0]
+                        return generate_simple_audio(text, speaker_id)
+                    
+                    simple_btn.click(
+                        simple_generate_with_mapping,
+                        inputs=[simple_text, simple_speaker],
+                        outputs=[simple_audio, simple_status]
+                    )
+                
+                # Context Generation
+                with gr.TabItem("🎭 Context Generation"):
+                    gr.Markdown("### Generate speech with audio context")
+                    
+                    with gr.Row():
+                        with gr.Column():
+                            context_text = gr.Textbox(
+                                label="Context Text",
+                                placeholder="މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ",
+                                value="އައިލީޝް އޭނާއަތަށް ލިބުނީ އެންމެ ނުވަ މަހުގަ",
                                 lines=2,
+                                elem_classes=["dhivehi-text"]
+                            )
+                            context_audio = gr.Audio(
+                                label="Context Audio",
+                                type="filepath"
+                            )
+                            target_text = gr.Textbox(
+                                label="Text to Generate",
                                 placeholder="މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ",
+                                value="މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ",
+                                lines=3,
                                 elem_classes=["dhivehi-text"]
                             )
+                            # Generate speaker choices for context generation
+                            context_speaker = gr.Radio(
+                                choices=[choice[0] for choice in speaker_choices],
+                                label="Speaker",
+                                value=speaker_choices[0][0] if speaker_choices else "0: Speaker"
+                            )
+                            context_btn = gr.Button("🎵 Generate with Context", variant="primary")
+                        
+                        with gr.Column():
+                            context_audio_out = gr.Audio(label="Generated Audio")
+                            context_status = gr.Textbox(label="Status", interactive=False)
+                    
+                    def context_generate_with_mapping(text, speaker_display, context_text_val, context_audio_val):
+                        # Extract speaker ID from display text
+                        speaker_id = speaker_display.split(":")[0]
+                        return generate_context_audio(text, speaker_id, context_text_val, context_audio_val)
+                    
+                    context_btn.click(
+                        context_generate_with_mapping,
+                        inputs=[target_text, context_speaker, context_text, context_audio],
+                        outputs=[context_audio_out, context_status]
+                    )
                 
-                with gr.Accordion("⚙️ Options", open=False):
-                    use_style = gr.Checkbox(
-                        label="Use audio style references",
-                        value=True
+                # Conversation Generation
+                with gr.TabItem("💬 Conversation"):
+                    gr.Markdown("### Generate dual-speaker conversations")
+                    
+                    with gr.Row():
+                        speaker_a = gr.Dropdown(
+                            choices=get_speaker_choices(),
+                            label="Speaker A",
+                            value=get_speaker_choices()[0]
+                        )
+                        speaker_b = gr.Dropdown(
+                            choices=get_speaker_choices(),
+                            label="Speaker B",
+                            value=get_speaker_choices()[1] if len(get_speaker_choices()) > 1 else get_speaker_choices()[0]
+                        )
+                    
+                    with gr.Accordion("🎵 Audio Style References", open=False):
+                        with gr.Row():
+                            with gr.Column():
+                                gr.Markdown("**Speaker A Context**")
+                                speaker_a_audio = gr.Audio(
+                                    type="filepath",
+                                    label="Speaker A Audio Style"
+                                )
+                                speaker_a_text = gr.Textbox(
+                                    label="Speaker A Context Text",
+                                    lines=2,
+                                    placeholder="މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ",
+                                    elem_classes=["dhivehi-text"]
+                                )
+                            
+                            with gr.Column():
+                                gr.Markdown("**Speaker B Context**")
+                                speaker_b_audio = gr.Audio(
+                                    type="filepath",
+                                    label="Speaker B Audio Style"
+                                )
+                                speaker_b_text = gr.Textbox(
+                                    label="Speaker B Context Text",
+                                    lines=2,
+                                    placeholder="މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ",
+                                    elem_classes=["dhivehi-text"]
+                                )
+                    
+                    with gr.Accordion("⚙️ Options", open=False):
+                        use_style = gr.Checkbox(
+                            label="Use audio style references",
+                            value=True
+                        )
+                        split_sentences_checkbox = gr.Checkbox(
+                            label="Split sentences",
+                            value=False
+                        )
+                    
+                    dialogue_text = gr.Textbox(
+                        lines=6,
+                        placeholder="ދަރިފުޅު މިއަދު ހާދަ ލަސްތިވީ.. މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ...... ކޮބާ ތިޔަ ރިޕޯޓް ފޮތް؟\nދަބަހުގަ އެބައޮތް!.\nވަދެބަލަ އެތެރެއަށް!... ދީބަލަ ރިޕޯޓްފޮތް މަންމަ ބަލައިލަން!",
+                        value="ދަރިފުޅު މިއަދު ހާދަ ލަސްތިވީ.. މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ...... ކޮބާ ތިޔަ ރިޕޯޓް ފޮތް؟\nދަބަހުގަ އެބައޮތް!.\nވަދެބަލަ އެތެރެއަށް!... ދީބަލަ ރިޕޯޓްފޮތް މަންމަ ބަލައިލަން!",
+                        label="Dialogue Lines (one per line)",
+                        elem_classes=["dhivehi-text"]
                     )
-                    split_sentences_checkbox = gr.Checkbox(
-                        label="Split sentences",
-                        value=False
+                    
+                    conv_btn = gr.Button("🎵 Generate Conversation", variant="primary")
+                    conv_audio = gr.Audio(label="Generated Conversation")
+                    conv_status = gr.Textbox(label="Status", interactive=False)
+                    
+                    conv_btn.click(
+                        generate_conversation,
+                        inputs=[speaker_a, speaker_b, speaker_a_audio, speaker_a_text,
+                            speaker_b_audio, speaker_b_text, dialogue_text,
+                            split_sentences_checkbox, use_style],
+                        outputs=[conv_audio, conv_status]
                     )
-                
-                dialogue_text = gr.Textbox(
-                    lines=6,
-                    placeholder="ދަރިފުޅު މިއަދު ހާދަ ލަސްތިވީ.. މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ...... ކޮބާ ތިޔަ ރިޕޯޓް ފޮތް؟\nދަބަހުގަ އެބައޮތް!.\nވަދެބަލަ އެތެރެއަށް!... ދީބަލަ ރިޕޯޓްފޮތް މަންމަ ބަލައިލަން!",
-                    value="ދަރިފުޅު މިއަދު ހާދަ ލަސްތިވީ.. މަންމަ ކިހާއިރެއް ދަރިފުޅުގެ އިންތިޒާރުގަ އިންނަތާ...... ކޮބާ ތިޔަ ރިޕޯޓް ފޮތް؟\nދަބަހުގަ އެބައޮތް!.\nވަދެބަލަ އެތެރެއަށް!... ދީބަލަ ރިޕޯޓްފޮތް މަންމަ ބަލައިލަން!",
-                    label="Dialogue Lines (one per line)",
-                    elem_classes=["dhivehi-text"]
-                )
-                
-                conv_btn = gr.Button("🎵 Generate Conversation", variant="primary")
-                conv_audio = gr.Audio(label="Generated Conversation")
-                conv_status = gr.Textbox(label="Status", interactive=False)
-                
-                conv_btn.click(
-                    generate_conversation,
-                    inputs=[speaker_a, speaker_b, speaker_a_audio, speaker_a_text,
-                           speaker_b_audio, speaker_b_text, dialogue_text,
-                           split_sentences_checkbox, use_style],
-                    outputs=[conv_audio, conv_status]
-                )
-        
-        # Wire up model change
-        model_dropdown.change(
-            change_model,
-            inputs=[model_dropdown],
-            outputs=[model_info, speaker_a, speaker_b, simple_speaker, context_speaker]
-        )
+            
+            # Wire up model change
+            model_dropdown.change(
+                change_model,
+                inputs=[model_dropdown],
+                outputs=[model_info, speaker_a, speaker_b, simple_speaker, context_speaker]
+            )
+            
+            gr.Markdown(f"""
+            ---
+            **Current Model:** {get_model_info()}
+            
+            **Tips:**
+            - Simple: Basic text-to-speech
+            - Context: Use reference audio for voice consistency
+            - Conversation: Multi-speaker dialogues with style control
+            
+            **Issues:**
+            - Context: Context breaks sometimes. Adding multiple context audio seems to make it work, or adding previous generation to the context helps.
+            - Audio: Sometimes the generated audio is not in sync with the text.
+            - Long sentences: Generated long sentences seems sped up.
+            - Repeating words: Generated text sometimes repeats words.
+            """)
         
-        gr.Markdown(f"""
-        ---
-        **Current Model:** {get_model_info()}
-        
-        **Tips:**
-        - Simple: Basic text-to-speech
-        - Context: Use reference audio for voice consistency
-        - Conversation: Multi-speaker dialogues with style control
-        
-        **Issues:**
-        - Context: Context breaks sometimes. Adding multiple context audio seems to make it work, or adding previous generation to the context helps.
-        - Audio: Sometimes the generated audio is not in sync with the text.
-        - Long sentences: Generated long sentences seems sped up.
-        - Repeating words: Generated text sometimes repeats words.
-        """)
-    
-    return app
-
 if __name__ == "__main__":
     # Initialize default model on startup
     initialize_default_model()
     
-    app = create_interface()
     app.launch(share=False)