State based sub-band LP Wiener filters for speech enhancement in car environments
ICASSP '00 Proceedings of the Acoustics, Speech, and Signal Processing, 2000. on IEEE International Conference - Volume 01
Stereo-based stochastic mapping for robust speech recognition
IEEE Transactions on Audio, Speech, and Language Processing
A variable step size LMS algorithm
IEEE Transactions on Signal Processing
A robust variable step-size LMS-type algorithm: analysis andsimulations
IEEE Transactions on Signal Processing
IEEE Transactions on Audio, Speech, and Language Processing
| Hi-index | 0.00 |
An investigation is made into whether non-acoustic noise reference signals can be used for noise estimation, and subsequently speech enhancement, in very low signal-to-noise ratio (SNR) environments where conventional noise estimation methods may be less effective. The environment selected is Formula 1 motor racing where SNRs fall frequently to -15dB. Analysis reveals three primary noise sources (engine, airflow and tyre) which are found to relate to data parameters measured by the car's onboard computer, namely engine speed, road speed and throttle opening. This leads to the proposal of a two stage noise reduction system that uses first engine speed to cancel engine noise within an adaptive filtering framework. Secondly, a maximum a posteriori (MAP) framework is developed to estimate airflow and tyre noise from data parameters which is subsequently removed. Objective measurements comparing noise estimation with conventional methods show the proposed method to be substantially more accurate. Subjective quality tests using comparative mean opinion score listening tests found that the proposed method achieves +1.43 compared to +0.66 for a conventional method. In subjective intelligibility tests, 81.8% of words were recognised correctly using the proposed method in comparison to 76.7% with no noise compensation and 66.0% for the conventional method.