The Potential of Speech as the Calibration Sound for Level Calibration of Non-Laboratory Listening Test Setups

Abstract

The pandemic of COVID-19 and the resulting countermeasures have made it difficult or impossible to perform listening tests in controlled laboratory environments. This paper examines the possibility of using speech for level calibration of sound reproduction systems used in listening tests performed in non-laboratory conditions, i.e., when such tests are distributed through the means of electronic communication and performed in a home environment. Moreover, a larger pool of potential test subjects can be reached in this manner. The perception of what the “normal” level of reproduced speech should be was examined through a listening experiment by letting the listeners set the level of reproduced speech samples as they saw fit, depending on the used sound reproduction system, the (non)existence of visual stimulus, and the voice of the speaker. The results show that the perception of normal speech level is highly individual when it comes to setting that level by listening to reproduced speech. The interindividual differences between the subjects are considerably larger than the impact of the three main effects. The understanding of what the “normal” level of read speech should be was examined experimentally as well by asking the subjects to read a paragraph of text, depending on the visual stimulus. The results show that the “normal” level of read speech is reasonably consistent and averages at 55 dBA at a normal conversational distance of 1 m, in a room with room acoustics conditions typical for home environment and low background noise, and with the visual stimulus that mimics the interlocutor put within the personal space of the reader. A preliminary proposal is given of a level calibration method for non-laboratory listening experiments based on these results, and some of its aspects that require further research are discussed.

Development and verification of non-supervised smartphone-based methods for assessing pure-tone thresholds and loudness perception

“…How to treat the uncalibrated mobile device and additional hardware in non-laboratory setups remains a challenge. Kisić et al (2022), for instance, proposed that human speech might be an appropriate and stable test signal for microphone calibration while Scharf et al (2023) considered the whistling sound of a 0.33 l beer bottle as a rough calibration signal.…”

Detecting Lombard Speech Using Deep Learning Approach

“…In reality, the speech signal is way more complicated than just simple Lombard vs. non-Lombard differentiation. One may even argue about the definition of “normal” speech [ 22 ]. For example, silence or unvoiced fragments may occur, or there may be a mixture of speech, etc.…”

“…For example, silence or unvoiced fragments may occur, or there may be a mixture of speech, etc. Moreover, rarely is it known whether recordings are collected in a controlled environment [ 22 ]. Therefore, the detection is always an approximation of the speech type; however, incorporating an averaging procedure allows for building a near real time Lombard speech detection process.…”

“…Various researchers have analyzed Lombard speech produced in different types and levels of noise for speech intelligibility [ 10 , 11 , 12 ], audio and audio-visual speech recognition [ 13 , 14 , 15 , 16 ], speaker recognition [ 17 , 18 , 19 ], and emotional speech analysis [ 20 ]. Overall, an automatic speech recognition system (ASR) performance may be degraded when Lombard speech is present in the speech signal [ 15 , 16 , 21 , 22 , 23 , 24 ]. This refers to the case when the ASR system is not trained on Lombard samples but only on “normal” (neutral) utterances.…”