A short-time objective intelligibility measure for time-frequency weighted noisy speech

Abstract

Existing objective speech-intelligibility measures are suitable for several types of degradation, however, it turns out that they are less appropriate for methods where noisy speech is processed by a timefrequency (TF) weighting, e.g., noise reduction and speech separation. In this paper, we present an objective intelligibility measure, which shows high correlation (rho=0.95) with the intelligibility of both noisy, and TF-weighted noisy speech. The proposed method shows significantly better performance than three other, more sophisticated, objective measures. Furthermore, it is based on an intermediate intelligibility measure for short-time (approximately 400 ms) TF-regions, and uses a simple DFT-based TF-decomposition. In addition, a free Matlab implementation is provided.Index Terms-intelligibility prediction, speech enhancement, noisy speech.

Signal Subspace Speech Enhancement with Oblique Projection and Normalization

“…Short-time objective intelligibility measure (STOI) is based on mean cross-correlations between processed and reference signals across time-frequency cells [30]. The STOI values for NB and WB speech are shown in Fig.…”

“…5 (a) and (b) for NB and WB speech respectively. computed as in (30), where S x ( j, m) andSx ( j, m) denote the spectral slopes of the clean and enhanced signals respectively of the j th band, m th frame and W ( j, m) represents the weight [29]. …”

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

“…The STOI [38], [40] method achieved a Spearman correlation coefficient of 0.94 with subjective word intelligibility scores validating its use for the ADN-TN and additive noise partition of the C-Qual database. The speech intelligibility estimation experiments are restricted to the additive noise partitions of the database for which we aim to predict the output of the intrusive STOI method (which is well correlated with inteligibility) using our non-intrusive estimator.…”

“…LCIA differs from the LCQA method [22] by employing an additional feature (iSNR), an external VAD, the use of a two-step feature selection and projection technique and training on databases labeled with STOI [38]. The LCIA method begins by deriving per frame features from the speech waveform, then applying a statistical model followed by a two-step dimensionality reduction and GMM mapping.…”

“…The SII score is monotonically related to the intelligibility of speech and is given in the range 0 to 1 (where a score of 0.5 means that half of the speech cues are audible and usable to the listener) [37]. More recently, the Short-Time Objective Intelligibility (STOI) method for intrusive intelligibility assessment has been proposed which has been shown to have a high correlation (greater than 0.92) with subjective intelligibility scores for noisy and noisesuppressed speech [38], [39], [40]. 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].…”