Finding quality podcasts is hard. Right now, the process is based on tactile and visual feedback. However, that’s not how you consume podcasts in the real world. 48% of podcasts are consumed at home, when people do something else. Podcasting is a passive experience, so you can't afford to be tied to your device.
Similarly, we realized that finding good information is also hard. Googling has become a game of avoiding SEO trolls and while content is plentiful, it's getting increasingly time intensive to distinguish between high quality sources and ad traps. We built Ava hoping to solve this problem by allowing you to get information and satisfy your curiosity by learning from the smartest people in the world.
Special thanks to all our friends and family who listened to us asking them about their podcasting habits during the past week or so. Our ideas are directly inspired by their needs and goals.
Why we think this could turn into a business?
First of all, the podcasting market is huge and growing at an unprecedented rate. 65% of active podcast listeners have started listening in the past year and 1/4 of Americans listen to a podcast weekly. Furthermore, on average, Americans listen to 7 podcasts weekly (~ 6 hours and 37 minutes time spent on average).
Furthermore, according to A16z, "the ad revenue is estimated to hit $500M in 2019" and has been doubling yearly.
Given this market, we believe we can monetize through referrals to podcast creators, advertising, and premium subscriptions. We would talk to users to better understand how to tackle product-market fit.
- Figures from the A16z Podcast Report: https://a16z.com/2019/05/23/podcast-ecosystem-investing-2019/
What it does
Ava answers your questions using a deep learning model that queries podcast transcripts. Using this personalized source of information, it's able to provide great answers and also suggest great content.
Why this team?
We're all from vastly different backgrounds, hailing from Canada, Romania, and the U.S. But what united us to come together for Ask Ava was our passion for empowering content creators using creative technical solutions. Between the three of us, we've seen the unique challenges of the audio medium while working at a radio station, worked hard to sell photo prints despite the steep costs of e-commerce platforms, and struggled with the question of how to promote indie content on YouTube. We hope that Ask Ava is the tool that will both be a boon for content creators and for budding podcast listeners.
How we built it
We built Ava using Alexa Skills and a server system that queries a BERT NLP deep learning model. The model is trained on the SQuAD data set from Stanford. We transcribed tens of hours of podcasts in order to feed the model so that we could then create a Q&A system based on podcasts.
The Alexa skill pings the server with a question and the server returns a response and a link to a podcast to play.
Challenges we ran into
Transcribing data was more difficult than we expected. Even transcribing a short podcast takes a lot of time, which meant that we had to be very efficient about what podcasts we chose.
Getting the model to run was also difficult, since we relied on IoT, ML, a web app client, multiple databases, a web scraper.
Accomplishments that I'm proud of
We made it work without having prior experience with AWS, EC2, S3, Alexa Skills, and web scraping. We've learned a lot in the past 24 hours alone.
We have four repos (client/server, web scraper, Alexa skill, and ML) which can be found here: yc-hacks
What's next for Ask Ava
We only had a limited amount of time to train our model on podcast transcripts so look forward to further training on new podcasts and refining our training methodology.