app.py

import os
from openai import OpenAI
import tiktoken

import nltk
nltk.download('punkt')
from nltk.tokenize import sent_tokenize

from pypdf import PdfReader, PageRange

## Function to read the uploaded PDF
def read_data_from_PDF(input_path):
  input_text = ''
  print ('Reading PDF from path', input_path)
  reader = PdfReader(input_path)
  number_of_pages = len(reader.pages)
  print ('PDF has been read with ', number_of_pages, ' pages')
  for page in reader.pages:
    input_text += page.extract_text() + "\n"
  return input_text


## Function to split the text into sentences
def split_text (input_text):
  split_texts = sent_tokenize(input_text)
  return split_texts


## Function to create chunks while considering sentences
def create_chunks(split_sents, max_token_len=50):
  enc = tiktoken.encoding_for_model("gpt-4o-mini")
  current_token_len = 0
  input_chunks = []
  current_chunk = ""
  for sents in split_sents:
    sent_token_len = len(enc.encode(sents))
    if (current_token_len + sent_token_len) > max_token_len:
      input_chunks.append(current_chunk)
      current_chunk = ""
      current_token_len = 0
    current_chunk = current_chunk + sents
    current_token_len = current_token_len + sent_token_len
  if current_chunk != "":
    input_chunks.append(current_chunk)
  return input_chunks


## Function to create chunks
def create_input_chunks(input_text):
  split_sents = split_text(input_text)
  input_chunks = create_chunks(split_sents, max_token_len=3000)
  return input_chunks


## Function to create summary of the given input text
def create_summary_points(input_chunks):
  client = OpenAI(api_key=userdata.get('OPENAI_API_KEY'))
  instructPrompt = """
                  You are provided with a piece of text regarding a subject. Please identify the key facts and highlights about this piece of text.
                  Please provide the output in the form of a list something like -

                  - Highlight 1
                  - Highlight 2
                  - Highlight 3
                  """
  podcastFacts = []
  for text in input_chunks:
    request = instructPrompt + '\n' + text
    chatOutput = client.chat.completions.create(model="gpt-4o-mini",
                                            messages=[{"role": "system", "content": "You are a helpful assistant."},
                                                      {"role": "user", "content": request}
                                                      ]
                                            )
    podcastFacts.append(chatOutput.choices[0].message.content)
  return "\n".join(podcastFacts)


## Two different prompt styles for the podcast conversation
debate_podcast_prompt = """
Could you simulate a podcast conversation in a debate-style between two experts, \"Sid\" and \"Darren\", discussing the following key points extracted from a research paper?
Some things that you need to keep in mind while creating the conversation:
- In the debate, Sid takes a stance that has a positive view of the findings and supports the implications and findings represented by these key points. They provide their reasoning and analogical examples to back up their interpretations.
- Conversely, Darren adopts a more critical or alternative viewpoint. They question some of the findings by discussing potential drawbacks, limitations, or different outcomes.
- The conversation should see both experts engaging with each key point, presenting their views, challenging each other's interpretations, and discussing the broader implications of their arguments.
- The debate should be balanced, allowing each expert to articulate their perspective comprehensively.
- Conclude the conversation with each expert summarizing their overall position on the topic.
Here's some of the facts from the topic.
"""

casual_podcast_prompt = """
Could you simulate a podcast conversation between \"Sid\" and \"Darren\" having a conversation about the following facts?
Some things I'd like to ask:
  - Use \"Sid:\" and \"Darren:\" to indicate who is speaking.
  - Start the dialog with a casual discussion on what each person is drinking right now.
  - Make the dialog about this as long as possible and make it sound funny
  - Sid is the one presenting the information, Darren is asking intelligent questions that help Sid elaborate the facts.
Here's some of the facts from the topic.
"""

styles = {'casual':casual_podcast_prompt,
          'debate': debate_podcast_prompt}


## Function to create the podcast script
def create_podcast_script(podcast_points, output_style):
  client = OpenAI(api_key=userdata.get('OPENAI_API_KEY'))
  instructPrompt = styles[output_style]
  request = instructPrompt + '\n' + podcast_points
  chatOutput = client.chat.completions.create(model="gpt-4o-mini",
                                            messages=[{"role": "system", "content": "You are a helpful assistant."},
                                                      {"role": "user", "content": request}
                                                      ]
                                            )
  return chatOutput.choices[0].message.content


## Function to call all the podcast script generation steps
def create_podcast(input_path, output_style):
  input_text = read_data_from_PDF(input_path)
  input_chunks = create_input_chunks(input_text)
  podcastHighlights = create_summary_points(input_chunks)
  podcastScript = create_podcast_script(podcastHighlights, output_style)
  return podcastScript


## Function to generate speech from input text
def openai_generation(input_text, speaker_voice, model_choice="tts-1"):
  client = OpenAI(api_key=userdata.get('OPENAI_API_KEY'))
  response = client.audio.speech.create(
      model=model_choice,
      voice=speaker_voice,
      input=input_text
  )
  return response.read()


## Function to generate complete audio podcast from script
## NOTE: this function assumes that there are only two speakers; please modify if you have multiple speakers in the script
def create_podcast_audio(podcastScript, speakerName1="Sid", speakerChoice1='onyx', speakerName2="Darren", speakerChoice2='echo'):
  genPodcast = []
  podcastLines = podcastScript.split('\n\n')
  podcastLineNumber = 0
  for line in podcastLines:
    if podcastLineNumber % 2 == 0:
      speakerChoice = speakerChoice1
      line = line.replace(speakerName1+":", '')
    else:
      speakerChoice = speakerChoice2
      line = line.replace(speakerName2+":", '')
    genVoice = openai_generation(input_text=line, speaker_voice=speakerChoice, model_choice="tts-1")
    genPodcast.append(genVoice)
    podcastLineNumber += 1
  with open("genPodcast.mp3", "wb") as f:
    for pod in genPodcast:
      f.write(pod)
  return "genPodcast.mp3"

import gradio as gr

def upload_file(file):
    return file.name

with gr.Blocks() as demo:
    file_output = gr.File()
    upload_button = gr.UploadButton("Click to Upload a PDF", file_types=[".pdf"], file_count="single")
    upload_button.upload(upload_file, upload_button, file_output)
    podcast_style = gr.Dropdown(styles.keys(), label="podcast_style")
    generate_podcast_button = gr.Button("Generate Podcast Script")
    podcast_script = gr.Textbox(interactive=True, label="podcast_script")

    generate_podcast_button.click(fn=create_podcast, inputs=[file_output, podcast_style], outputs=podcast_script, api_name="generate_podcast_script")

    generate_audio_button = gr.Button("Generate Audio Version")
    podcast_audio = gr.Audio(label="podcast_audio", interactive=False, type="filepath")
    generate_audio_button.click(fn=create_podcast_audio, inputs=podcast_script, outputs=podcast_audio, api_name="generate_podcast_audio")

demo.launch(debug=True, share=True)