How to Create an AI-Generated News Podcast
The idea of an AI-generated podcast sounds unworkable, but is it? A video by Eric Shanfelt, publishing consultant and owner of Nearview Media, shows how easy it can be to deploy. Note: To copy a tested script using this strategy, go to the Media Prompt Library here.
Challenge
Shanfelt started experimenting with text-to-speech by automating every article for a few publications, much like the Washington Post and Atlantic Magazine have done.
However, he found that text-to-speech for every article had a low usage rate and became expensive.
A news round-up podcast model solves these issues. It can serve as an audio email newsletter, offering a streamlined method for followers to stay informed daily or weekly. While speech-to-text made news slow to absorb, a summarised audio report is be easier and faster. The audience can listen to a three-minute take on the top stories while driving or over coffee, i.e., doing something else.
Shanfelt asked the right question, “What if we could publish a podcast of all our stories, maybe every day, maybe every week, and have it all be completely done via AI? What if it could all be completely automated?”
In the long run, podcasts present a new revenue opportunity, but the first challenge was to develop a dedicated following by producing a daily or weekly podcast, manageable by a small team with other job responsibilities.
Strategy
Shanfelt created a daily news round-up podcast for Providence Business News using a few steps ( video by Eric Shanfelt is here).
- Upgrade to ChatGPT 4
ChatGPT 3.5 cannot read a URL, so you have to cut and paste the stories. ChatGPT 4, at $20 a month, has a LinkReader, which can reach the content inside any URL, so no cutting and pasting is required.
Keep in mind that ChatGPT3.5 also has a limit of processing 2006 tokens, which is about 1300 to 1500 words per prompt if you cut and paste.
2. Create the prompt
Shanfelt starts by finding RSS feed – any URL followed by /feed to the URL, of the news sources. So if the site is PBN.com, the RSS feed is PBN.com/feed.
The feed usually includes the title and publication date of any article, a critical input for GPT to create the script for a timely news round-up.
He then asked ChatGPT to create the podcast script, including several directives:
*Use the URL of the RSS feed
*Create a podcast script
*With an intro and outro
*With a {number of } sentence summary of {number of or all} articles
*With a two-sentence summary of each article
*From this {date} to {date}
*Without unspoken information, such as section heads
(See a wholly composed prompt in the Media Prompt Library here).
He used a two-sentence summary for the PBN daily podcast. An early version delivered production notes such as “Music up” and other sound effects; for example, he asked it to be kept as a script that the host would read without any production notes.
“So chat GPT got all excited and said, “Of course. Here’s the revised script without any production notes.”
GPT4 delivered:
“Welcome to the Providence Business News Daily Roundup. Today, we’re diving into the top stories from Rhode Island and southeastern Massachusetts. First up Rhode Island free clinic welcomes new medical director…”
When Shanfelt saw it, he was elated. “I’m like, boom and cha-ching. This is what I’m looking for. This is exactly the way I would read it out.”
He now puts the instructions into a single query.
3. Using a Text-to-Speech Platform
Shanfelt then dropped the script into a text-to-speech platform. He tested Amazon Polly (pay as you go, or $4 per 1 million characters) but settled on Beyond Words, which is 10,000 free for the first project, then $89 a month. Beyond Words can split the transcript into two voices – Shanfelt selected one male and one female – to make the audio more interesting. It also has a more friendly user interface than Polly, so it is easier to deploy in seconds or hand off to an admin. The voice also has a slightly more human-like experience, including some added tone, emotion, and natural pauses.
Several other platforms have their own unique properties. Descript is more potent for advanced publishers who want more editing capabilities. It offers Overdub, allowing users to synthesize their voice and various editing tools for webinars and vodcasts. CereProc, another option, has a “voice shop” and claims to offer the most nuanced emotional speech synthesis for those who want the most life-like voice possible.
They all have embeddable players. However, for most publishers in the NPB community, BeyondWords has everything you will need.
4. Upload the script to a Podcast hosting platform
Shanfelt recommends Buzzsprout. It is a super user-friendly and robust hosting platform that distributes podcasts across 100+ outlets, including the largest ones, Apple, Google, and Spotify. Its website has an extensive “learn to” library and email customer support, which it claims has an average wait time of 26 minutes rather than 24 hours, which has become standard for most email support services.
Conclusions and lessons learned
- Looking towards the future, Shanfelt says he is in the process of fully automating PBN’s daily podcast
- Some newsletters that curate stories from other sources can use this workflow to create an innovative, fully automated podcast.
- Prompt engineering is not like Google search. While the results are spectacular, it does require some time spent experimenting with the prompts and query results in a single conversation.
- The best is yet to come. Fully automated integrations are increasingly available.
