Kryptonite researchers have developed a new language identification (LID) classifier designed to optimize audio processing workflows. The team aimed to create a solution that maintains high recognition quality while significantly reducing computational overhead for large-scale speech recognition tasks.
The project team, led by Alexander Samoylov with contributions from Alexander Tararin and Artem Ryzhenkov, utilized a baseline model capable of distinguishing 126 languages. This original model contained approximately 1 billion parameters, which the team sought to optimize for faster performance and smaller footprint.
By applying knowledge distillation, the researchers trained a smaller model using the larger network as a teacher. The resulting classifier is approximately 10 times faster and 5 times lighter than the original, while also demonstrating increased resilience to audio distortions.
This development allows for more efficient scaling of automatic speech recognition (ASR) systems on existing infrastructure. By identifying the language before routing audio to specific recognition engines, the company can process large volumes of data with fewer resources.
