Working streaming and cool looking app

app.py

@@ -7,12 +7,14 @@ import soundfile as sf
 import gradio as gr
 import numpy as np
 import time
+import pymupdf
+import requests
 
 import torch
 from huggingface_hub import InferenceClient
 from kokoro import KModel, KPipeline
 # -----------------------------------------------------------------------------
-# Get podcast subject
+# Get default podcast materials from Daily papers
 # -----------------------------------------------------------------------------
 from papers import PaperManager
 
@@ -33,12 +35,12 @@ client = InferenceClient(
 )
 
 
-def generate_podcast_text(subject: str, steering_question: str | None = None) -> str:
+def generate_podcast_script(subject: str, steering_question: str | None = None) -> str:
     """Ask the LLM for a script of a podcast given by two hosts."""
     messages = [
         {"role": "system", "content": SYSTEM_PROMPT},
         {"role": "user", "content": f"""Here is the topic: it's the top trending paper on Hugging Face daily papers today. You will need to analyze it by bringing profound insights.
-{subject[:1000]}"""},
+{subject[:10000]}"""},
     ]
     if steering_question and len(steering_question) > 0:
         messages.append({"role": "user", "content": f"You could focus on this question: {steering_question}"})
@@ -58,7 +60,7 @@ def generate_podcast_text(subject: str, steering_question: str | None = None) ->
 # -----------------------------------------------------------------------------
 CUDA_AVAILABLE = torch.cuda.is_available()
 
-kmodel = KModel().to("cuda" if CUDA_AVAILABLE else "cpu").eval()
+kmodel = KModel(repo_id='hexgrad/Kokoro-82M').to("cuda" if CUDA_AVAILABLE else "cpu").eval()
 kpipeline = KPipeline(lang_code="a")  # English voices
 
 MALE_VOICE = "am_michael"  # [MIKE]
@@ -68,14 +70,23 @@ FEMALE_VOICE = "af_heart"  # [JANE]
 for v in (MALE_VOICE, FEMALE_VOICE):
     kpipeline.load_voice(v)
 
-# -----------------------------------------------------------------------------
-# Audio generation system with queue
-# -----------------------------------------------------------------------------
-
 @spaces.GPU
-def generate_podcast(pdf, url, topic):
-    podcast_text = generate_podcast_text(PODCAST_SUBJECT, topic)
-    lines = [l for l in podcast_text.strip().splitlines() if l.strip()]
+def generate_podcast(url: str, pdf_path: str, topic: str):
+    if pdf_path:
+        with pymupdf.open(pdf_path) as pdf_doc:
+            material_text = ""
+            for page in pdf_doc:
+                material_text += page.get_text()
+    elif url:
+        response = requests.get(f'https://r.jina.ai/{url}')
+        material_text = response.text
+    else:
+        material_text = PODCAST_SUBJECT
+    
+    # Generate podcast script!
+    podcast_script = generate_podcast_script(material_text, topic)
+
+    lines = [l for l in podcast_script.strip().splitlines() if l.strip()]
 
     pipeline = kpipeline
     pipeline_voice_female = pipeline.load_voice(FEMALE_VOICE)
@@ -114,28 +125,22 @@ demo = gr.Interface(
 If you do not specify any source materials below, the podcast will be about the top trending [Daily paper](https://huggingface.co/papers/), '**{list(top_papers.keys())[0]}**'""",
     fn=generate_podcast,
     inputs=[
+        gr.Textbox(
+            label="🔗 Type a Webpage URL to discuss this page (Optional)",
+        ),
         gr.File(
-            label="Optional - Upload a pdf",
+            label="📄 Upload a PDF to use it as discussion material (Optional)",
             file_types=[".pdf"],
             file_count="single",
         ),
-        gr.Textbox(
-            label="Optional - Type a URL to read its page",
-        ),
-        gr.Textbox(label="Do you have a more specific topic or question on the materials?"),
-        # gr.Dropdown(
-        #     label=UI_INPUTS["length"]["label"],
-        #     choices=UI_INPUTS["length"]["choices"],
-        #     value=UI_INPUTS["length"]["value"],
-        # ),
+        gr.Textbox(label="🤔 Do you have a more specific topic or question on the materials?", placeholder="You can leave this blank."),
     ],
     outputs=[
         gr.Audio(
-            label="Listen to your podcast",
+            label="Listen to your podcast 🔊",
             format="wav",
             streaming=True,
         ),
-        # gr.Markdown(label=UI_OUTPUTS["transcript"]["label"]),
     ],
     theme=gr.themes.Soft(),
     submit_btn="Generate podcast 🎙️", 

prompts.py

@@ -46,10 +46,8 @@ Ensure the dialogue has a natural ebb and flow:
 
 IMPORTANT RULE: Each line of dialogue should go in a new line [JANE] or [MIKE], as follows:
 
-[JANE] Hello, how are you?
-[MIKE] I'm good, thank you. How are you?
-[JANE] I'm good, thank you.
-[MIKE] Great.
+[JANE] Hello Mike, how are you?
+[MIKE] Nice to see you again, Jane. I'm very good. Today's topic is fascinating, because...
 
-Remember: Each intervention from a host should be on the same line.
+Remember: Each turn from a host should be on the same line.
 """