Evaluation of short-time speech-based intelligibility metrics

The Speech Transmission Index (STI) is based on acoustic measurements in environments and has been shown to be correlated with speech intelligibility under a wide range of acoustic conditions (Houtgast & Steeneken 1984). It is a weighted average of metrics derived from envelope signals in multiple frequency bands spanning the speech spectrum. A variety of methods have been proposed to compute the STI (Houtgast & Steeneken 1971; Steeneken & Houtgast 1980; Ludvigsen 1987; Drullman et al. 1994a, b; Payton et al. 1994; Drullman 1995; IEC 1998; Payton & Braida 1999; Payton et al. 2002; Goldsworthy & Greenberg 2004). Some of these methods use speech as the test stimulus rather than artificially modulated noise as originally proposed by Houtgast and Steeneken (1985). Many of the speech-based techniques have been shown to provide the same result as the traditional STI (Ludvigsen et al. 1990; Payton et al. 2002), which is based on modulation reductions in intensitymodulated noise and as a theoretically derived STI which is obtained from weighted signal-to-noise ratios (SNRs) in seven octave bands and room reverberation time (RT) (Houtgast & Steeneken 1985). To date, all speech-based approaches have used speech materials lasting at least a minute or two to generate metrics correlated with long-term speech intelligibility. Consequently, they have not been used to predict short-time changes in intelligibility due to time-varying environments such as fluctuating background noise. The current work investigates the ability of two speech-based methods to track short-term STI results by using speech segments of various lengths to compute results for environments with stationary speech-shaped noise, speechshaped noise plus reverberation or multi-talker babble. The methods that will be evaluated are the Envelope Regression (ER) and the Normalized Correlation (NC) methods. The ER method is based on the speech-based STI method proposed by Ludvigsen et al. (1990). The NC method was proposed by Goldsworthy and Greenberg (2004) who also analyzed the long-term characteristics of both metrics.