AI generated podcasts – are they any good for meeting preparation, or for conference interpreter training? My student Rebecca Bremer, a real podcast aficionado, wanted to find out and did some testing. Here’s what she has found out:

 Guest article by Rebecca Bremer, student of conference interpreting at TH Köln

General considerations

Podcasts have become a popular medium for learning, entertainment, and professional development. With advancements in AI, tools like NotebookLM and Jellypod now enable the creation of entire podcasts without human intervention. But how effective are AI-generated podcasts for studying, particularly for aspiring interpreters? This article explores their benefits, limitations, and overall usefulness.

Podcast-generating AI technology: What is it?

Creating a podcast traditionally involves a lot of work: research, writing a script, recording each episode, and editing. AI reduces this process by generating, curating, and producing content with minimal human input. Users only need to select a topic, prompt the AI, and customise details such as format and voice.

Why use podcasts for (interpreting) studies?

Benefits

• Emotional connection and contextualisation: Listening to podcasts we develop an emotional connection to them. This emotional connection aids with retaining the knowledge gained whilst listening.
  At the same time, the context within which the content is delivered makes it easier to understand.

• Flexibility and mobility: Podcasts can be accessed anytime, making them ideal for on-the-go learning whether it is during the day or night, on the way to class or an interpreting job.

• Auditory comprehension: Listening to podcasts enhances listening and terminological comprehension and exposes us to various dialects and accents.

• Targeted training: AI-generated podcasts, in particular, allow users to match content, speed, terminology and language to their needs. This can also be achieved with human-made podcasts. However, it requires more research and effort.

Limitations

• Data security when dealing with AI: Who has access to my data when I upload it to the AI platform? Where is it stored and how will it be used?

• Content: AI may produce repetitive, monotonous content with inaccuracies.

• Language: Many AI tools primarily support English, limiting accessibility for other languages.

• Ethical concerns: AI usage raises issues such as its environmental impact and job displacement.

Find out more about misleading AI generated content and AI’s environmental impact compiled by MIT university news.

The tools and what they do

NotebookLM

This AI program is part of the Google Labs Project Portfolio and was originally meant to be used as an AI notebook for structuring and summarising uploaded data. However, the program gained popularity for its audio summary feature. With this feature all uploaded information is summed up and an audio file of two hosts discussing the topics created.
The set-up of NotebookLM consists of three columns:

1. Column: Uploading source material

1. Column: A chat for AI prompting

1. Column: A studio for organising the uploaded material, which includes the audio summary feature

NotebookLM only offers audio summaries in English.

Jellypod

The program Jellypod is based on NotebookLM, but focused on podcast creation. Jellypod is meant for creating an entire, long-term podcast rather than just one episode. Its set-up also looks fairly simple but includes more steps than NotebookLM.
First one needs to create and generate the hosts they want to use in their podcasts. One can chose from a huge database of hosts or create a new one using their own voice file or generate one with AI.
As of January, Jellypod languages other than English for creating podcasts include German, Spanish, Hindi and Arabic.

Testing the tools

In the following section, both programs will be analysed based on factors such as voice quality, content accuracy, terminology use, and efficiency. The same source material—a presentation on British and German education systems—was used for both AI tools. It was based on a presentation held within the context of an interpreting class on multilingual conferences. Because of this context there is also a student glossary available to aid with analysing the terminology used in both AI podcasts. This topic was chosen due to being commonly well-known and, therefore, being able to assume that the AI has already been trained with different data on said topic.

This is what a NotebookLM podcast sounds like:

To listen to my Jellypod-generated podcast, click here.

Voice and speech

• NotebookLM produced natural-sounding conversations with varied intonation, making the listening experience more engaging. The voices of the hosts cannot be changed and only English is available.

• Jellypod offered a wider range of voices and accents ranging from the Queen’s English to rural Chilean Spanish. However, it sounded flat and mechanical, reducing engagement.

Content

Both podcasts maintained logical structures but omitted specific details, such as distinctions between British public and state schools. Nor was any information derived from the pictures given in the source, which made up an important part of the presentation. Additional information was used minimally and mostly to naturally connect one topic with the next one.

Terminology

While terminology was correctly used and all the education-related terminology used also appeared in the source material, only one-third of the terms included in the student glossary appeared in either podcast.
Pronunciation of foreign terminology was inconsistent with both AI tools.

Efficiency

A small test group (four participants) assessed knowledge retention using AI-generated podcasts versus a traditional presentation. Those who studied with podcasts scored lower (with the NotebookLM listener scoring slightly better results than the Jellypod listener), suggesting that AI-generated content may not be the most effective standalone study tool. However, effectiveness depends on individual learning styles. In addition, further tests with a bigger test group would need to be carried out to get more representative results.

Time-saving potential

• NotebookLM: Podcast generation took approximately six minutes. Because of its simple set-up, no additional time is needed for preparatory set-up steps such as the ones offered by Jellypod.

• Jellypod: Generation took a little bit longer than four minutes. These four minutes do not include preparatory set-up steps such as creating or choosing the hosts.

If a relevant podcast is already available on streaming platforms, it may be quicker to find one than to generate it. This depends on the topic of the podcast.

Conclusion

The question of how valuable using podcast generating AI is for interpreting (studies) does not have a definitive answer.
AI-generated podcasts offer valuable tools for interpreters, providing flexible and targeted study techniques. However, limitations in accuracy, engagement, and terminology completeness suggest they should be used as a supplementary resource rather than a primary study method. When using AI-generated content, it is crucial to verify accuracy, consider ethical implications, and reflect on one’s learning preferences. While promising, current AI tools like NotebookLM and Jellypod are still evolving and require further refinement to become truly effective study aids.
Personally, I prefer to use NotebookLM due to the more authentic and pleasant listening experience. However, there are no indicators that the NotebookLM AI will be adding other languages to its repertoire in the near future. For this reason, Jellypod would be preferrable for interpreters/interpreting students working with other languages.