app.py

import queue
import threading
import os
import gradio as gr
from dia.model import Dia
from huggingface_hub import InferenceClient
import numpy as np

# Hardcoded podcast subject
PODCAST_SUBJECT = "The future of AI and its impact on society"

# Initialize the inference client
client = InferenceClient("Qwen/Qwen2.5-Coder-32B-Instruct", provider="together", token=os.getenv("HF_TOKEN"))
model = Dia.from_pretrained("nari-labs/Dia-1.6B", compute_dtype="float16")

# Queue for audio streaming
audio_queue = queue.Queue()
stop_signal = threading.Event()


def generate_podcast_text(subject):
    prompt = f"""Generate a podcast told by 2 hosts about {subject}.
The podcast should be an insightful discussion, with some amount of playful banter.
Separate dialog as follows using [S1] for the male host and [S2] for the female host, for instance:
[S1] Hello, how are you?
[S2] I'm good, thank you. How are you?
[S1] I'm good, thank you. (laughs)
[S2] Great.
Now go on, make 5 minutes of podcast.
"""
    response = client.chat_completion([{"role": "user", "content": prompt}], max_tokens=1000)
    return response.choices[0].message.content


def split_podcast_into_chunks(podcast_text, chunk_size=3):
    lines = podcast_text.strip().split("\n")
    chunks = []

    for i in range(0, len(lines), chunk_size):
        chunk = "\n".join(lines[i : i + chunk_size])
        chunks.append(chunk)

    return chunks


def process_audio_chunks(podcast_text):
    chunks = split_podcast_into_chunks(podcast_text)
    sample_rate = 36000 # Modified from https://huggingface.co/spaces/nari-labs/Dia-1.6B/blob/main/app.py has 44100
    for chunk in chunks:
        if stop_signal.is_set():
            break

        raw_audio = model.generate(chunk, use_torch_compile=True, verbose=False)
        audio_chunk = np.array(raw_audio, dtype=np.float32)
        audio_queue.put((sample_rate, audio_chunk))

    audio_queue.put(None)


def stream_audio_generator(podcast_text):
    """Creates a generator that yields audio chunks for streaming"""
    stop_signal.clear()

    # Start audio generation in a separate thread
    gen_thread = threading.Thread(target=process_audio_chunks, args=(podcast_text,))
    gen_thread.start()

    try:
        while True:
            # Get next chunk from queue
            chunk = audio_queue.get()

            # None signals end of generation
            if chunk is None:
                break

            # Yield the audio chunk with sample rate
            yield chunk

    except Exception as e:
        print(f"Error in streaming: {e}")


def stop_generation():
    stop_signal.set()
    return "Generation stopped"


def generate_podcast():
    podcast_text = generate_podcast_text(PODCAST_SUBJECT)
    return podcast_text


with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("# NotebookLM Podcast Generator")

    with gr.Row():
        with gr.Column(scale=2):
            gr.Markdown(f"## Current Topic: {PODCAST_SUBJECT}")
            gr.Markdown("This app generates a podcast discussion between two hosts about the specified topic.")

            generate_btn = gr.Button("Generate Podcast Script", variant="primary")
            podcast_output = gr.Textbox(label="Generated Podcast Script", lines=15)

            gr.Markdown("## Audio Preview")
            gr.Markdown("Click below to hear the podcast with realistic voices:")

            with gr.Row():
                start_audio_btn = gr.Button("▶️ Generate Podcast", variant="secondary")
                stop_btn = gr.Button("⏹️ Stop", variant="stop")

            audio_output = gr.Audio(label="Podcast Audio", streaming=True)
            status_text = gr.Textbox(label="Status", visible=True)

    generate_btn.click(fn=generate_podcast, outputs=podcast_output)

    start_audio_btn.click(fn=stream_audio_generator, inputs=podcast_output, outputs=audio_output)
    stop_btn.click(fn=stop_generation, outputs=status_text)

if __name__ == "__main__":
    demo.queue().launch()