Low-complexity F0-based speech/nonspeech discrimination approach for digital hearing aids

Voicing detection based on adaptive aperiodicity thresholding for speech enhancement in non‐stationary noise

“…These state probability equations were chosen experimentally based on the observation of experimental data. Observe that, rather than estimating the likelihoods of the states using Gaussian mixture models (which require training the parameters of the Gaussians), likelihoods are here directly computed by (14)- (16) from the values of ap 0 (t), α(t), r n (t, 0) and r s (t, 0). No other features are extracted and no training is performed.…”

“…No other features are extracted and no training is performed. The model receives as input data (observation) the sequence of probabilities in (14)- (16)…”

“…(4) Determine α(t) using (10). (5) Compute the probabilities in (14)- (16) and decide on the current voicing state using the Viterbi algorithm. (6) Return r x (t, 0),r ′ n (t, 0), F 0 (t) and the voicing decision.…”

“…This approximation was shown to achieve good performance when applied to sound classification in hearing aids. In this work, we extend the experimental evaluation given in [16] by implementing this efficient difference function within the proposed voicing detector. The results demonstrate that this efficient implementation remains effective for voicing detection and speech enhancement, showing only a moderate degradation in comparison with the standard difference function.…”

“…Generally, this kind of device imposes severe constraints in terms of computational capacity, which avoid the implementation of common signal processing algorithms, including the conventional difference function. In a previous work [16], we proposed an algorithm to compute an accurate approximation of the difference function which reduces drastically the number of required operations. This approximation was shown to achieve good performance when applied to sound classification in hearing aids.…”