PMC Corporation primarily sells a store visit management system that uses vehicle license plate recognition. We spoke with Nobutaka Otsuka, section chief of the IT Solutions Division's System Development Department, about the background and effects of adopting the AmiVoice API.

Background and Challenges in Using Speech Recognition

Limited ability to check conversations between customers and support center

We sell visitor management systems that use vehicle license plate recognition for commercial facilities such as car dealerships, as well as entrance/exit management systems for factories and parking lots that use license plate recognition, and entrance/exit management systems that use QR codes and RFID (which reads data wirelessly using radio waves).

Our support center receives inquiries from store staff at car dealerships that use our system, but because the industry does not allow the use of other means of communication such as chat tools or email forms, communication over the phone is extremely important.
When the number of incoming calls was low, we listened to the recordings to check whether there were any problems with how we responded to customers (such as using prohibited words or whether the conversation was going smoothly), but as the number of incoming calls increased, we were unable to keep up with the work of listening to the recordings, and we were no longer able to check all of them. Therefore, we decided to convert the calls into text and check the recorded data in order to reduce the amount of time spent on each call.

Key Drivers of the Adoption

Accurate recognition even with intonation differences and fillers

When introducing speech recognition, we compared it with other companies' services, but what we felt was overwhelmingly superior about AmiVoice was its tolerance for fluctuations in intonation and accent, and the fact that fillers (words like "ah" and "ere") were properly recognized as fillers, without being combined with other words to create a separate word. We also like the convenient word registration function, which allows us to register our company's service name and terms specific to the automotive industry, so that the words are properly displayed when transcribed. We also use the "speaker diarization" function, which identifies speakers in multi-person conversations, and the "sentiment analysis" function, which can determine emotions from voice.

When it came to implementation, the provided sample program was well-commented, so I was able to implement it using only that. For some features that were not provided as samples, such as the optional sentiment analysis function, I was able to implement it using only the official documentation on the website. The sample source was very concise and easy to understand, so I was able to implement it without any problems overall. With other companies' services, there are often cases where the software does not work properly due to different versions, or the error messages differ from the documentation, but I was very grateful that this was not the case with this service.

Impact of the Implementation and Future Outlook

Text conversion of calls allows for optimization of human resources

The support center handles around 10 to 20 calls per day, each lasting around 5 to 20 minutes. Calls between the support center and customers are recorded and stored on a cloud PBX, and the recorded data is converted into text and stored as data in an in-house CRM.
Previously, we had to assign dedicated staff to listen to and check the content of phone calls, but by converting the content into text, we can use it to analyze prohibited words and as a source for conversation summaries using ChatGPT. By automating this process, we can now allocate the time that dedicated staff would have spent on other tasks, thereby optimizing our human resources.