Pythian created a data model for Sonos that would predict when a network problem would disrupt audio quality, reducing support calls for this problem by 60%. This system utilized Hadoop and Oracle database technologies.


Imagine listening to your favorite song, only to have it pause unexpectedly. With more and more wireless devices fighting for bandwidth in the home, Sonos wanted to take proactive steps to ensure the company continued to deliver on its promise of Rock Solid Wireless. They turned to Pythian to design a system and data model that would predict when a network problem would cause a drop in audio quality so Sonos could proactively provide support to their customers.


Sonos is the leading manufacturer of wireless music systems. The smart speaker system lets listeners stream all the music on earth wirelessly, in any room of the home with control from a wifi-enabled device. The system enables listeners to play not only their personal digital music collection, but gives them access to millions of songs and thousands of radio stations by partnering with the leading online music streaming services.


Critical Issues

The inherent unpredictability of home internet networks, as well as the ambient environment in which their speakers live, means that there are instances in which customers will experience drops in audio connections, degrading the listening experience. Imagine listening to your favourite song only to have it pause unexpectedly; we’ve all experienced it -- that annoying buffering while the network tries to resolve problems and re-establish a connection with its host.

If Sonos could predict the conditions under which a given device will have a drop in audio quality, solutions could be developed that would guide the user through troubleshooting steps to resolve the most common issues on their own—for example, rebooting the router or changing wireless channel. The Sonos user experience would be improved, and calls to the support center would be reduced, significantly lowering overall support costs while at the same time improving customer satisfaction.

Sonos wanted to know if they could predict audio drop offs by examining various Wi-Fi metrics collected from the Sonos devices. Sonos turned to Pythian’s data science team to discover whether the diagnostic data Sonos was collecting from the devices included the right data to build the predictive model.

We Provided

Pythian developed a proof-of-concept (PoC) modeling environment to assess the data. After analyzing and experimenting with dozens of metrics to determine which ones were the best predictors of potential audio drop-offs, the team selected the best features for use in the predictive model.

The Sonos data was stored in a cloud environment, so Pythian worked closely with the cloud provider to fully leverage the machine learning features to develop the model.  A data pipeline was created and transformed using Hive, and finally saved to blob storage. The predictive pipeline was then built in the cloud fetching data directly from the blob storage.


Pythian built a set of PoC data models using anonymized diagnostic data from connected Sonos devices that can predict drop-outs at an accuracy rate of 56%, while keeping the false positive rate at just 6%. Catching these drop-offs, which are typically caused by busy or crowded home networks, before they happen can ensure an epic listening experience for Sonos owners, and up to 60% fewer calls to the support centre for this issue.

The Pythian advanced analytics team developed a model using diagnostic data from connected Sonos devices such as packet send attempts, physical errors and latency.


Various system diagnostics and Wi-Fi metrics were used to develop a data model that would predict when a given Sonos system is likely to experience a drop in audio quality as a result of network problems upstream, so that Sonos could proactively provide customers with a diagnosis and solution to the problem.

  • Hadoop

  • Oracle Databases