Update app.py

app.py

@@ -1,7 +1,4 @@
 import gradio as gr
-import google.generativeai as genai
-import numpy as np
-import re
 import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
 from huggingface_hub import snapshot_download, login
@@ -10,9 +7,7 @@ import os
 import spaces
 import warnings
 from snac import SNAC
-from dotenv import load_dotenv
-
-load_dotenv()
+import numpy as np
 
 logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
 logger = logging.getLogger(__name__)
@@ -27,6 +22,8 @@ model = None
 tokenizer = None
 snac_model = None
 
+EMOTIVE_TAGS = ["<laugh>", "<sigh>", "<gasp>", "<cry>", "<yawn>"]
+
 @spaces.GPU()
 def load_model():
     global model, tokenizer, snac_model
@@ -59,40 +56,6 @@ def load_model():
         logger.error(f"Error loading model: {str(e)}")
         raise
 
-@spaces.GPU()
-def generate_podcast_script(api_key, content, uploaded_file, duration, num_hosts):
-    try:
-        genai.configure(api_key=api_key)
-        model = genai.GenerativeModel('gemini-2.5-pro-preview-03-25')
-        
-        combined_content = content or ""
-        if uploaded_file:
-            file_content = uploaded_file.read().decode('utf-8')
-            combined_content += "\n" + file_content if combined_content else file_content
-        
-        prompt = f"""
-        Create a podcast script for {'one person' if num_hosts == 1 else 'two people'} discussing:
-        {combined_content}
-        
-        Duration: {duration}. Include natural speech, humor, and occasional off-topic thoughts.
-        Use speech fillers like um, ah. Vary emotional tone.
-        
-        Format: {'Monologue' if num_hosts == 1 else 'Alternating dialogue'} without speaker labels.
-        Separate {'paragraphs' if num_hosts == 1 else 'lines'} with blank lines.
-        
-        Use emotion tags in angle brackets: <laugh>, <sigh>, <chuckle>, <cough>, <sniffle>, <groan>, <yawn>, <gasp>.
-        
-        Example: "I can't believe I stayed up all night <yawn> only to find out the meeting was canceled <groan>."
-        
-        Ensure content flows naturally and stays on topic. Match the script length to {duration}.
-        """
-        
-        response = model.generate_content(prompt)
-        return re.sub(r'[^a-zA-Z0-9\s.,?!<>]', '', response.text)
-    except Exception as e:
-        logger.error(f"Error generating podcast script: {str(e)}")
-        raise
-
 def process_prompt(prompt, voice, tokenizer, device):
     prompt = f"{voice}: {prompt}"
     input_ids = tokenizer(prompt, return_tensors="pt").input_ids
@@ -155,17 +118,15 @@ def redistribute_codes(code_list, snac_model):
     return audio_hat.detach().squeeze().cpu().numpy()
 
 @spaces.GPU()
-def text_to_speech(text, voice, temperature=0.6, top_p=0.95, repetition_penalty=1.1, max_new_tokens=1200):
-    global model, tokenizer, snac_model
-    if model is None or tokenizer is None or snac_model is None:
-        load_model()
-    
+def generate_speech(text, voice, temperature, top_p, repetition_penalty, max_new_tokens, progress=gr.Progress()):
     if not text.strip():
         return None
     
     try:
+        progress(0.1, "Processing text...")
         input_ids, attention_mask = process_prompt(text, voice, tokenizer, device)
         
+        progress(0.3, "Generating speech tokens...")
         with torch.no_grad():
             generated_ids = model.generate(
                 input_ids=input_ids,
@@ -179,76 +140,75 @@ def text_to_speech(text, voice, temperature=0.6, top_p=0.95, repetition_penalty=
                 eos_token_id=128258,
             )
         
+        progress(0.6, "Processing speech tokens...")
         code_list = parse_output(generated_ids)
+        
+        progress(0.8, "Converting to audio...")
         audio_samples = redistribute_codes(code_list, snac_model)
         
         return (24000, audio_samples)
     except Exception as e:
-        logger.error(f"Error in text_to_speech: {str(e)}")
-        raise
-
-@spaces.GPU()
-def render_podcast(api_key, script, voice1, voice2, num_hosts):
-    try:
-        lines = [line for line in script.split('\n') if line.strip()]
-        audio_segments = []
-
-        for i, line in enumerate(lines):
-            voice = voice1 if num_hosts == 1 or i % 2 == 0 else voice2
-            try:
-                result = text_to_speech(line, voice)
-                if result is not None:
-                    sample_rate, audio = result
-                    audio_segments.append(audio)
-            except Exception as e:
-                logger.error(f"Error processing audio segment: {str(e)}")
-
-        if not audio_segments:
-            logger.warning("No valid audio segments were generated.")
-            return (24000, np.zeros(24000, dtype=np.float32))
-
-        podcast_audio = np.concatenate(audio_segments)
-        podcast_audio = np.clip(podcast_audio, -1, 1)
-        podcast_audio = (podcast_audio * 32767).astype(np.int16)
-        
-        return (24000, podcast_audio)
-    except Exception as e:
-        logger.error(f"Error rendering podcast: {str(e)}")
-        raise
+        print(f"Error generating speech: {e}")
+        return None
 
-with gr.Blocks() as demo:
-    gr.Markdown("# AI Podcast Generator")
+with gr.Blocks(title="Orpheus Text-to-Speech") as demo:
+    gr.Markdown(f"""
+    # 🎵 [Orpheus Text-to-Speech](https://github.com/canopyai/Orpheus-TTS)
+    Enter your text below and hear it converted to natural-sounding speech with the Orpheus TTS model.
     
-    api_key_input = gr.Textbox(label="Enter your Gemini API Key", type="password")
+    ## Tips for better prompts:
+    - Add paralinguistic elements like {", ".join(EMOTIVE_TAGS)} or `uhm` for more human-like speech.
+    - Longer text prompts generally work better than very short phrases
+    """)
     
     with gr.Row():
-        content_input = gr.Textbox( label="Paste your content (optional)", lines=8 )
-        document_upload = gr.File(label="Upload Document (optional)")
-    
-    duration = gr.Radio(["1-5 min", "5-10 min", "10-15 min"], label="Estimated podcast duration", value="1-5 min")
-    num_hosts = gr.Radio([1, 2], label="Number of podcast hosts", value=2)
-    
-    voice_options = ["tara", "leah", "jess", "leo", "dan", "mia", "zac", "zoe"]
-    voice1_select = gr.Dropdown(label="Select Voice 1", choices=voice_options, value="tara")
-    voice2_select = gr.Dropdown(label="Select Voice 2", choices=voice_options, value="leo")
-    
-    generate_btn = gr.Button("Generate Script")
-    script_output = gr.Textbox(label="Generated Script", lines=10)
-    
-    render_btn = gr.Button("Render Podcast")
-    audio_output = gr.Audio(label="Generated Podcast")
+        with gr.Column():
+            text_input = gr.Textbox(
+                label="Text Input",
+                placeholder="Enter the text you want to convert to speech...",
+                lines=8
+            )
+            voice_select = gr.Dropdown(
+                choices=["tara", "leah", "jess", "leo", "dan", "mia", "zac", "zoe"],
+                value="tara",
+                label="Voice"
+            )
+            with gr.Accordion("Advanced Options", open=False):
+                temperature = gr.Slider(
+                    minimum=0.1, maximum=1.0, value=0.6, step=0.1,
+                    label="Temperature",
+                    info="Higher values increase randomness in the output"
+                )
+                top_p = gr.Slider(
+                    minimum=0.1, maximum=1.0, value=0.95, step=0.05,
+                    label="Top-p",
+                    info="Lower values increase determinism in the output"
+                )
+                repetition_penalty = gr.Slider(
+                    minimum=1.0, maximum=2.0, value=1.1, step=0.1,
+                    label="Repetition Penalty",
+                    info="Higher values discourage repetitive patterns"
+                )
+                max_new_tokens = gr.Slider(
+                    minimum=100, maximum=2000, value=1200, step=100,
+                    label="Max Length", 
+                    info="Maximum length of generated audio (in tokens)"
+                )
+            
+            with gr.Row():
+                submit_btn = gr.Button("Generate Speech", variant="primary")
+                clear_btn = gr.Button("Clear")
+        
+        with gr.Column():
+            audio_output = gr.Audio(label="Generated Speech")
     
-    generate_btn.click(generate_podcast_script, 
-                       inputs=[api_key_input, content_input, document_upload, duration, num_hosts], 
-                       outputs=script_output)
+    submit_btn.click(
+        generate_speech,
+        inputs=[text_input, voice_select, temperature, top_p, repetition_penalty, max_new_tokens],
+        outputs=audio_output
+    )
     
-    render_btn.click(render_podcast, 
-                     inputs=[api_key_input, script_output, voice1_select, voice2_select, num_hosts], 
-                     outputs=audio_output)
-
-    num_hosts.change(lambda x: gr.update(visible=x == 2), 
-                     inputs=[num_hosts], 
-                     outputs=[voice2_select])
+    clear_btn.click(lambda: "", inputs=None, outputs=text_input)
 
 if __name__ == "__main__":
     try: