A robust 400-bps speech coder against background noise

The 400-bps speech was unintelligible in helicopter noise. A detailed study has shown that both voicing errors and spectral distortion are the key to the loss of intelligibility. The voicing accuracy is restored by first generating robust features whose distributions are not strongly dependent upon noise levels. These features are then appropriately combined using linear programming. A better spectrum is generated by using the approach of adaptive vector quantization. This means the codebook is updated based upon noise level and subsequently searched to determine the best clean codeword for synthesis. Both approaches show promise for improved speech intelligibility in noise for a low-rate coder