I write a newsletter called Above Average where I talk about the second order insights behind everything that is happening in big tech. If you are in tech and don’t want to be average, subscribe to it.
A lot of people want podcasts transcribed and read instead of listen to them. We can go one level up and even extract insights from the podcasts as well using Open AI API. Here’s my tweet exchange which provoked this experiment.
So here is what we are going to do. And it will slightly different from what I suggested in the tweet. The goal is to pick a YouTube video and get transcription of that video and then using prompt engineering we extract insights, ideas, book quotes & summary etc.,
To Summarize we achieve our goal in three steps:
- Step 1: Select a YouTube podcast video
- Step 2: Transcribe the video
- Step 3: Get Insights from the transcription
Step 1: Select a YouTube podcast video
A recent podcast conversation that broke YouTube was Jeff Bezos on Lex Friedman podcast. So for this exercise I will pick this video.
Step 2: Transcribe the video
I used langchain YouTubeAudioLoader along with Open AI’s audio to text model whisper to transcribe the youtube video. As usual you would need your Open AI secret key to use the following script.
import os
import sys
import openai
from dotenv import load_dotenv, find_dotenv
_ = load_dotenv(find_dotenv()) # read local .env file
openai.api_key = os.environ['OPENAI_API_KEY']
## youtube video's audio loader - langchain
from langchain_community.document_loaders.blob_loaders.youtube_audio import YoutubeAudioLoader
from langchain_community.document_loaders.generic import GenericLoader
from langchain_community.document_loaders.parsers import OpenAIWhisperParser #, OpenAIWhisperParserLocal
url="https://www.youtube.com/watch?v=DcWqzZ3I2cY&ab_channel=LexFridman"
save_dir="outputs/youtube/"
loader = GenericLoader(
YoutubeAudioLoader([url],save_dir),
OpenAIWhisperParser()
)
docs = loader.load()
print(docs[0].page_content[0:500])
# Specify the file path where you want to save the text
file_path = "audio-transcript.txt"
try:
with open(file_path, 'a', encoding='utf-8') as file:
for doc in docs:
file.write(doc.page_content)
print(f'Large text saved to {file_path}')
except FileNotFoundError:
print(f"Error: Input file '{file_path}' not found.")
except Exception as e:
print(f"An error occurred: {e}")
You might see the following error during running this script and I pasted the solution that works in case you are using a Windows system.
- ERROR: Postprocessing: ffprobe and ffmpeg not found. Please install or provide the path using –ffmpeg-location.
Running this script will generate the transcript and store it in a text file audio-transcript.txt.
Step 3: Extract insights from the conversation
To extract insights, I am using Open AI API and here is the script. The code is loading the transcript text and passing it along with a prompt designed to extract insights, people & books. To get more interesting thigs out of this conversation you can come up with a more interesting prompt. Note that the file name is slightly different because I had to cut the transcript to a short length since my completion query to Open AI API was exceeding my TPM limits.
import os
import sys
import openai
import shutil
from pprint import pprint
from dotenv import load_dotenv, find_dotenv
_ = load_dotenv(find_dotenv()) # read local .env file
openai.api_key = os.environ['OPENAI_API_KEY']
client = openai.OpenAI()
file_path = "audio-transcript-copy.txt"
try:
with open(file_path, 'r', encoding='utf-8') as file:
long_text = file.read()
print(f'{file_path} is rad')
except FileNotFoundError:
print(f"Error: Input file '{file_path}' not found.")
except Exception as e:
print(f"An error occurred: {e}")
prompt2 = f"""
You will be provided with text deilimited by triple quotes.
The given text is a podcast transcript.
Provide the host and guest name.
Summarize the transcript in to 10 points.
If there are any people referred in the transcript. Extract the people mentioned and list them along with some info about them in the following format
1. Person 1's Name: Person 1's profession or what he or she is known for or the context in which he or she was referred to.
2. Person 2's Name: Person 2's profession or what he or she is known for or the context in which he or she was referred to.
...
2. Person N's Name: Person N's profession or what he or she is known for or the context in which he or she was referred to.
If the transcript doesnt contain refereces to any people then simply write \"No people referred to in the conversation.\"
Extract the books mentioned and list them in the following format.
1. Book 1's Title: Context in which the book was referred to.
2. Book 2's Title: Context in which the book was referred to.
...
N. Book N's Title: Context in which the book was referred to.
If the transcript doesnt contain refereces to any books then simply write \"No books referred to in the conversation.\"
IF you find any inspiration quotoes complie them in to a list.
\"\"\"{long_text}\"\"\"
"""
response = client.chat.completions.create(
model="gpt-4",
messages=[
{
"role": "user",
"content": prompt2
}
],
temperature=0.7,
#max_tokens=64,
#top_p=1
)
print(response.choices[0].message.content)
Here is what the output I got:
AI PRODUCT IDEA ALERT 1: Can provide a service to podcasters to generate smart transcripts with insights. This would be a B2B play.
AI PRODUCT IDEA ALERT 2: Instead of a service to podcast creators, it could be a B2C customers who listen to podcasts and want to get read through podcasts and want to create their own library of insights.
Expect both these ideas to be used by existing podcast hosting companies like Spotify & launch these ideas as new features. If I was a Product Manager in any such companies I would be pitching them by now.
That’s it for day 5 of 100 Days of AI.
Follow me on Twitter, LinkedIn for latest updates on 100 days of AI or bookmark this page.