A Single-Channel Non-Intrusive C50 Estimator Correlated With Speech Recognition Performance

Abstract

Abstract-Several intrusive measures of reverberation can be computed from measured and simulated room impulse responses, over the full frequency band or for each individual mel-frequency subband. It is initially shown that full-band clarity index C50 is the most correlated measure on average with reverberant speech recognition performance. This corroborates previous findings but now for the dataset to be used in this study. We extend the previous findings to show that C50 also exhibits the highest mutual information on average. Motivated by these extended findings, a non-intrusive room acoustic (NIRA) estimation method is proposed to estimate C50 from only the reverberant speech signal. The NIRA method is a data-driven approach based on computing a number of features from the speech signal and it employs these features to train a model used to perform the estimation. The choice of features and learning techniques are explored in this work using an evaluation set which comprises approximately 100000 different reverberant signals (around 93 hours of speech) including reverberation from measured and simulated room impulse responses. The feature importance of each feature with respect to the estimation of the target C50 is analysed following two different approaches. In both cases the newly chosen set of features shows high importance for the target. The best C50 estimator provides a root mean square deviation around 3 dB on average for all reverberant test environments.

A data-driven non-intrusive measure of speech quality and intelligibility

“…In this paper we limit our discussion to the case of additive noise and the effects of speech codecs. Other degradations such as reverberation and effects of noise suppression are not considered here but the reader is referred to the appropriate literature elsewhere on nonintrusive reverberation estimation such as [43]. The remainder of this paper is organized as follows, in Section II we review the non-intrusive LCQA method [22] and describe LCIA, a development of LCQA for non-intrusive speech intelligiblity assessment.…”

“…A non-intrusive method for speech quality and intelligibility estimation of reverberant and de-reverberated speech was presented in [41] and a method for predicting the intelligibility of reverberant speech for cochlear implant listeners was presented in [42]. A recent method for non-intrusive estimation of the C50 room acoustic parameter was presented in [43] and shown to correlate with speech quality and speech recognition performance. The low complexity speech intelligibility algorithm (LCIA) is a datadriven non-intrusive measure that has been shown to have a high per-condition correlation with subjective intelligibility scores of noisy and noise suppressed speech [44].…”

Predicting Word Error Rate for Reverberant Speech

“…Tsilfidis et al showed that C50 and D50 can be used to predict phoneme recognition rate [15]. Parada et al proposed a blind C50 estimator that is correlated with phoneme recognition [16].…”

An image-inspired audio sharpness index

“…However, in many cases, the clean signal is not available (e.g., audio record or voice received in telephony), so that non-intrusive (or reference-free) measures are required. Most of them, like NIRA [6] or NISI [7], are based on machinelearning techniques and derive indicators from a large set of signal parameters by maximizing the correlation with reference indicators on a training corpus. The drawback of this approach is that the indicators depend on the training conditions and that they are blind to the physical grounds of intelligibility.…”