Effects of signal bandwidth and noise on individual speaker identification

Abstract

Two experiments were conducted to evaluate the effects of increasing spectral bandwidth from 3 to 10 kHz on individual speaker recognition in noisy conditions (+5, 0, and −5 dB signal-to-noise ratio). Experiment 1 utilized h(Vowel)d (hVd) signals, while experiment 2 utilized sentences from the Rainbow Passage. Both experiments showed significant improvements in individual speaker identification in the 10 kHz bandwidth condition (6% for hVds; 10% for sentences). These results coincide with the extant machine recognition literature demonstrating significant amounts of individual speaker information present in the speech signal above approximately 3–4 kHz. Cues from the high-frequency region for speaker identity warrant further study.

Classification of indexical and segmental features of human speech using low- and high-frequency energy

“…(3) Given the findings of this study and previous behavioral findings demonstrating the benefits of extending the bandwidth from 3 to 10 kHz for identifying a talker of interest in noise (Schwartz et al, 2018), it is important to consider how speech processing algorithms designed to amplify the high-frequency region of speech (e.g., frequency-lowering techniques) potentially alter natural acoustic cues regarding talker identity [for a review of frequency-lowering techniques, refer to Alexander (2013)]. In noisy conditions where low frequencies are significantly masked by environmental noise, the information above 3-4 kHz may increase in importance.…”

“…Segmental information related to phoneme and sentence recognition (e.g., Donai and Paschall, 2015;Vitela et al, 2015) as well as indexical features of the talker, such as sex (e.g., Monson et al, 2014;Donai and Lass, 2015;Donai and Halbritter, 2017) and talker identity (Schwartz et al, 2018), have also been found in the high-frequency region of the speech signal. Some studies limited auditory access only to high-frequency energy (HFE) above 3-5 kHz (e.g., Donai and Paschall, 2015;Donai and Halbritter, 2017;Monson et al, 2014;Vitela et al, 2015), while others varied the spectral bandwidth of the stimuli to purposely include or exclude HFE above 8 kHz (e.g., Schwartz et al, 2018, Flaherty et al, 2021Polspoel et al, 2022).…”