Artificial Intelligence has been the buzzword for the better first half of 2023 with more and more people taking notice. Leading the discussion and general increase in the awareness of AI has been OpenAI’s ChatGPT.

We’ve seen that what started out as a general interest and the occasional coffee break conversation starter amongst the team soon became some of the team adding AI tools like ChatGPT and Stable Diffusion into their daily workflow. From helping to make refinements and tweaks to marketing messages to using image generation models to help with design and branding visualisation mock-ups.

We’ve seen firsthand how AI has helped us with our day-to-day tasks and as a business that develops products to help our clients and users achieve a goal, we believe AI has the potential to deliver better experiences and more value to the end user.

With this belief in mind, this year, at Digital Workroom some of our team have been experimenting with how we can meaningfully add AI to our products. One example of this was upgrading the dictation/transcription capabilities of our app Noted.

The Process

To create a better product we first looked at all the ways AI could bring value to our users, eventually we settled on enhanced transcription via AI.

Noted’s transcription feature was at first built around a general-purpose speech recognition model which was released by OpenAI in 2022, the model was trained on a large dataset of diverse audio. However, it was soon found that full integration and direct usage of the model into the App was not possible as there were some issues:

1. Data Security and Privacy

The original model from OpenAI required users to upload their data to the cloud, which means users would have to share their data with OpenAI. Noted is a tool for people to help store anything from general day-to-day memos to important and potentially confidential meeting information. With this in mind, our product has always had an emphasis on user data security and privacy. We also wanted this transcription feature to be on device with no reliance on the user needing an internet connection. Therefore, the use of the OpenAI model was not possible. However, during R&D we were able to explore the option of using custom projects and ultimately ended up utilising a custom model that was derived from the original OpenAI model mentioned above. This custom model fit our criteria of being able to run independently on the user's device without an internet connection which meant all data would stay within the user's device.

2. Performance and Portability

Having found a core model we could integrate into our app the next step was optimising the feature with an importance placed particularly on speed and accuracy. After much testing, the team settled on the idea that we offer different levels of transcription for users to choose from. The lighter levels would be based on smaller versions of the model which would take up less space on the user’s device and offer the advantage of faster operation times. The tradeoff in using the smaller models would be in some of the transcriptions' accuracy. On the other hand, higher levels of transcription would mean using larger versions of the model which would take up more space, be harder to run on devices due to requiring more system resources, and be generally slower. However, the level of accuracy and detail in the produced transcriptions was incredibly impressive and we were sure users who went with this option would be pleasantly surprised.

Summary

The rollout of Noted's first AI feature went smoothly with the app receiving a great amount of positive feedback from our users, many of which were glowing reviews about the new feature and its performance.

The team is now hard at work with researching ways to further improve the transcription feature. Currently, projects involving exploring newly updated technologies like CoreML and Metal are underway. We hope to integrate these projects and launch them in the near future, allowing users to experience an even better transcription service.

Overall, implementing AI into Noted has been a very successful project for the team, and going forward the team is incredibly excited to find new and interesting applications of AI and integrate them into our products as well as the products of our clients.