Introduction.
Nice to meet you. My name is Yoshinaga, and I am the developer of Interview AI.
I am the head of product development at Antaa, a platform service for physicians.
Usually, my job is to prioritize product development and hold development meetings with engineers, but the other day I had the opportunity to conduct a career interview with a teacher who uses Antaa (Antaa also runs a career business called “D35” and publishes interviews with teachers). (Antaa also has a career project called “D35,” which features interviews with teachers).
On September 9, 2024, we interviewed Dr. Masatsugu Hamaji, Chief of the Department of Pulmonary Surgery at Nara Medical University Hospital, for about an hour from 9:30 pm.
The interview was conducted online using the Zoom recording feature. The interview itself was very informative and enjoyable.
But after that, I had to transcribe the interview (roughly 10,000 words for one hour), revise it to a natural conversational style, and add titles and subheadings . I started the project thinking, “I wish I could use an existing transcription service and get it done quickly, but I wonder if it will take much time to rewrite.
Use existing transcription services, but…
I immediately used a transcription service, but it took some time for the transcription to be generated, and since I talk while thinking during the interview, many filler words such as “um,” “uh,” etc. were all transcribed.
So we had to do quite a bit of revision work, such as deleting filler words, taking into account the overall flow of the conversation, and significantly modifying the order, which took about 2-3 hours in total.
The interview article itself was released in three parts (Part 1, Part 2, and Part 3) two days later without incident, thanks to the very quick confirmation by Dr. Hamaji, who agreed to conduct the interview.
However, the editing process was quite difficult,
I wonder if we can manage this editing process. It’s going to be a lot of work to interview teachers and then transcribe them into articles, and even now, other employees are doing it, too.
I felt that this was the right time to start developing the interview AI (September 12, 2024).
Looking back, I had a hard time in the past.
Thinking back, when we launched Merp, a web-based medical interview service for medical institutions in 2016, we were also constantly conducting interviews for case studies.
Melp case study articles (more than 50)
In the same way, I would make an appointment for an interview online, do a 30-minute Zoom recording of the interview (Zoom was not widely used at the time, so I had to start by explaining what Zoom was), and since there were no transcription tools with high accuracy yet, I would play the audio and stop, I had to play back the audio, stop, transcribe, and play back the audio again, which took even more time than before.
At the time, I was still searching for a convenient transcription tool, and tried using Google Docs’ transcription tool. I tried using the Google Docs transcription tool, but the accuracy of the transcription was still low, and I thought it would take more time to correct the transcription, so I ended up playing back the audio and transcribing it.
Focus in the development of interview AI
Therefore, in the hope of solving the above issues, we focused on the following three points in the development of Interview AI.
It took me about two weeks to develop the robot, because I also had my day job and Mia (a talking cat-shaped robot that speaks dialect), but I personally think that I was satisfied with the product, which I would want to use myself.
If we have the opportunity to conduct interviews in the future, I think we can use the interview AI so that the whole process will take less than 15 minutes for a one-hour interview, when creating three articles of 3,000 words each.
If you are experiencing challenges in transcribing interviews or dialogues, we encourage you to try it.
