A novel training method to separate speech from unwanted background noise.

The Need

Separation of speech from background noise is critical to many listening devices. To generalize these devices to new speakers and background noises in classification based speech separation algorithms, the classifiers need to be trained on a large number of acoustic conditions, which is a big challenge compared with previous methods. Therefore, it is critical to develop a large-scale training method for listening devices.

The Technology

Researchers at The Ohio State University, led by Dr. DeLiang Wang, developed a novel large-scale training method that helps listening devices separate speech from unwanted background noise. This invention formulates recognition of speech from background noise as a classification problem to simplify the separation issue to a case of binary classification. While previous art that employs Gaussian mixture models and support vector machines perform well in trained environments, these methods have high complexity and are difficult to generalize to new environments. In this invention, a new technique has been found for efficient large-scale training to overcome the generalization issue.

Commercial Applications

  • Mobile devices and headsets
  • Automatic speech and speech recognition
  • Listening devices
  • Microphones

Benefits/ Advantages

  • Scalability and flexibility – the system can be practically trained on a large number of acoustic conditions
  • Simultaneously addresses voiced and unvoiced speech separation from background interference
  • Outperforms existing speech separation methods
  • Generalizes well to a variety of new environments, including new speakers and new stationary and non-stationary noises
  • Separates noisy speech efficiently

Interested in this Technology?

Submit your interest below