A Comparison of Noise Reduction Techniques for Speech Recognition in Telecommunications Environments
AG Maher, RW King, JG Rathmell · 1992
The recognition performance of telecommunications applications of automatic speech recognition is likely to be degraded significantly by noise generated in the speaker's environment and by the telephone channel. This paper describes a study of techniques for noise reduction which can be applied at the input to standard recognizers trained on noise-free speech. Initial results, using signal-to-noise ratio measures, show that a spectral subtraction technique can provide up to 10 dB improvement in SNR, compared with 6 dB for an adaptive line enhancement method. Speech and pitch detection processes, which are required to support the noise reduction algorithms themselves, are also outlined.