Speech reinforcement in noisy reverberant environments using a perceptual distortion measure

Abstract

In this paper, a time-frequency weighting is proposed for speech reinforcement (near-end listening enhancement) in a noisy and reverberant environment, which optimizes a perceptual distortion measure locally for each time-frequency bin. The algorithm acts as a dynamic range compressor, smearing out the energy of the clean speech along time.Simulations predict an intelligibility increase with respect to the unprocessed condition and two reference methods, for moderate smoothing windows, as measured by the optimized distortion measure and two objective intelligibility measures.

Speech reinforcement with a globally optimized perceptual distortion measure for noisy reverberant channels

“…where D(f, t) is the local distortion in a single T/F bin [12], where we substitute A(f, t) = α 2 (f, t) with the implicit constraint A(f, t) > 0, and where we define…”

“…In this work, we will use several mathematical properties of (1) which were derived in [12]. For deterministic speech s, uncorrelated stochastic disturbance segments , 1,2 and scaling factors α, β, within T/F bin (f, t), we have…”

“…We follow the model of [12] for the received corrupted speech x, which is written down as (see also Fig. 1)…”

“…It was shown that, although the distortion measure was designed for assessing perceptual quality, it could also be used for intelligibility enhancement due to its short-time sensitivity. In [12], we extended the algorithm of [10] partially to include reverberant corruptions, where the optimization was performed locally for each time-frequency (T/F) bin independently. In this paper, we complete the extension of [10,12] for the noisy reverberant case, by globally optimizing distortion summed over a number of T/F bins.…”

“…In [12], we extended the algorithm of [10] partially to include reverberant corruptions, where the optimization was performed locally for each time-frequency (T/F) bin independently. In this paper, we complete the extension of [10,12] for the noisy reverberant case, by globally optimizing distortion summed over a number of T/F bins. We show that this global problem can be written down as a posynomial program, which can be elegantly solved in polynomial time by e.g., interior point methods [13].…”