| 摘要 |
<p>For treating a speech signal in order to attenuate the noise contribution before transmission, each frame of the sampled signal, digitised and formatted into frames of determined length, is subjected to a processing of the form: ¦Y[ omega ]¦ = ¦X[ omega ]¦-opt( alpha ( omega ))¦B[ omega ]¦ applied to the transform of each frame in the frequency domain, where: ¦X[ omega ]¦ is the estimation of the amplitude of the speech modified by the noise; ¦B[ omega ]¦ is an estimate of the amplitude of the noise, obtained from an observation of the noise, during silence periods of the speech, estimated using an estimator having a sufficient time constant; alpha ( omega ) is an overestimation factor of the noise, calculated over a predetermined number of noise frames upstream of the processed frame; opt[ alpha ( omega )] constitutes a weighting of the subtractive term, which is a function of the signal-to-noise ratio in a particular frequency band, this weighting being banded. Application to the recognition of speech using Hidden Markov Models. <IMAGE></p> |
