Tim Fingscheidt

Overcoming the statistical independence assumption w.r.t. frequency in speech enhancement

In this paper, we give a solution on how to overcome the assumption of statistical independence of adjacent frequency bins in noise reduction techniques. We show that under relaxed assumptions the problem results in an a-priori SNR estimation problem, where all available noisy speech spectral amplitudes (observations) are exploited. Any state-of-the-art noise power spectral density (psd) estimation and weighting rule can be used - they do not need to be restated. In order to solve for an estimator well suited for real-time applications, we model the a-priori SNR values as Markov processes w.r.t. frequency. On the basis of the formulation by Ephraim and Malah, this leads to a new a-priori SNR estimator that yields fewer musical tones.

Efficient coding of transitional speech segments in CELP

Speech coding at low and medium bit rates benefits from the assumption that the speech signal is stationary inside a subframe or even inside consecutive subframes. Speech signals however comprise a considerable amount of transient segments, such as onsets, which are usually difficult to encode with good quality. This paper proposes a method to improve the encoding of transient (sub-)frames by disregarding the adaptive codebook contribution and strengthening the fixed codebook contribution. This can principally be applied to any CELP-like standard speech coder while keeping the same bit rate. The proposed technique has been successfully employed in the ITU-T 4 kbps speech coder candidate commonly proposed by AT&T, Conexant currently Mindspeed, Deutsche Telekom, France Telecom, Matsushita, NTT, and Siemens in November 2001. Simulation results show the improvements achieved by this approach.