How can we assist Vietnamese users in listening to a broad range of written news articles while on the go?

CodeLink identified a gap in the market for voice-read news articles curated from all of the main Vietnamese news outlets. They wanted to apply their proprietary text-to-speech and speech-to-text AI model to solve this problem. To achieve this goal, CodeLink tasked their internal teams with building a mobile application that aggregates news from the top Vietnamese publishers and lets users listen to the curated articles.

The CodeLink internal team worked as a fully autonomous team to design and build out the Web application, iOS, and Android mobile applications.

The initial web release of the platform took 3 months, with the mobile applications developed over 6 weeks.

The team conducted research and evaluation of various Spectrogram models like Tacotron2 and FastPitch, as well as different Vocoder models like HifiGan. They then built a dataset of 10 hours of audio and used speech-to-text for auto annotation and fine-tuning with human supervision to label 10 hours of audio. After training the model from scratch using the home-grown dataset, the team deployed the inference cluster onto GCP and optimized it to save costs. Bao Noi is now fully released and available on iOS, Android, and the web, providing users with a seamless listening experience.