Effect of Type of Speech Equalization and Averaging Method on The Long-Term Average Speech Spectra of Five Indian Languages and British English

Purpose: To compare two methods for determining the long-term average speech spectrum (LTASS) across languages - root-mean-square equalization and averaging (RMSe) and loudness equalization and averaging (Le) - and to assess their implications for hearing-aid fitting methods. Method: Speech samples from multiple talkers of British English (BE), Indian English (IE), and four Indian languages were analyzed. For RMSe, the samples were equalized in overall RMS amplitude and the RMS amplitude in each 1/3 octave band was averaged across samples. For Le, the loudness of each talker’s speech was equated using a loudness model, and estimated loudness values in each 1/3 octave band were averaged. A preliminary experiment confirmed that equalizing the peak long-term loudness predicted by the model accurately equated speech loudness across languages. Results: With RMSe, LTASS differences across languages were observed primarily at high frequencies whereas with Le there were differences between BE and the Indian languages over a wide frequency range for both male and female speakers. Values of the speech intelligibility index (SII) for an LTASS input at 60 dB SPL were calculated for several hypothetical patterns of hearing loss after application of frequency-dependent gains intended to give good audibility for speech at moderate levels. For the Indian languages, the SII was consistently higher when Le was used to determine the LTASS than when RMSe was used to determine the LTASS. Conclusions: The LTASS determined using Le provides a more perceptually relevant characterization of cross-language spectral differences than LTASS determined using RMSe. Hearing-aid fitting methods may be improved using language-specific LTASS values derived using Le.