A single AI system improves song picks on smart speaker music service

Digital music services rely on software pipelines to choose which tracks play next. These systems often connect many separate pieces of software. One group selects possible songs, while other programs score and rank them.

A team of researchers tested a single model called Sona to handle all these tasks at once on Yandex Music. The project was described in a research paper submitted in August 2026 by Alexandr Udeneev and thirty-three other authors.

The old setup used more than fifteen separate tools to find song candidates. It then passed those tracks through pre-ranking and ranking steps that evaluated hundreds of different details, including data from a large model named Argus.

Sona replaced this entire chain with one model built around a shared profile of each listener. The design uses an encoder, which turns a history of user actions into machine data, alongside a decoder and a ranking module.

The model reads the timeline of past actions directly from event logs. It does not rely on hand-crafted rules or manually shaped features to understand what people enjoy hearing.

During training, a larger teacher model helps guide how tracks are scored. Once the system is finished and running live for users, that teacher model is removed, leaving only Sona to make choices.

The team tested Sona against the existing setup during an experiment on live traffic. The trial ran on My Vibe, a recommendation stream played through smart speakers.

The test showed clear gains over the old system. The number of active users rose by 4.53 percent, which was the main goal measured by the team.

Listeners also kept songs playing longer. Total listening time increased by 6.30 percent, and the number of likes went up by 11.42 percent.

The increase in active users was more than double the improvement delivered by Argus, which had previously been the top tool used on that music stream.