Task

To develop a real-time audio classifier capable of accurately classifying music, speech, and noise.

Challanges

• We had to integrate our solution within a complex existing Python codebase.
• The audio classifier needed to operate in real time, with minimal look-ahead available.

How we helped

• Our solution integrated seamlessly with Dolby.io's existing codebase.
• We created and trained a "best-in-class" neural network model based on TensorFlow and Python, achieving a high level of accuracy and recall despite real-time constraints.
• We helped create a new dataset for training audio classifiers in general.
• We developed a framework for assessing and comparing the performance of audio classifiers.
• We collaborated effectively with Dolby.io’s team and followed their working processes (Scrum agile development).

Results

We created a new production-ready audio classifier for the Dolby.io API suite.