Acoustic correlates of intelligibility enhancements in clearly produced fricatives

Kazumi Maniwa, Allard Jongman, Travis WadeInterspeech 2007

TL;DR

This study investigates how "clear speech" – speaking more deliberately – improves the intelligibility of fricative sounds, particularly for listeners with simulated hearing loss. They found that while clear speech generally helps, the acoustic features that contribute to this benefit differ between normal-hearing and hearing-impaired listeners, suggesting different strategies might be needed for speech enhancement technologies.

What Problem Does This Paper Address?

Fricative consonants (like 's', 'f', 'sh') are notoriously difficult to distinguish, especially for individuals with hearing impairments or in noisy environments. While "clear speech" (speaking more slowly and distinctly) is known to boost intelligibility, the specific acoustic mechanisms behind this enhancement, and whether they apply equally to listeners with different auditory capabilities, remain unclear. This paper aims to bridge this gap by examining how clear speech affects fricative intelligibility and identifying the acoustic correlates of these improvements, with a particular focus on listeners with simulated hearing loss.

Methods

The researchers conducted two experiments. In Experiment 1, they tested normal-hearing listeners on their ability to identify fricatives in minimal pairs (e.g., /s/ vs. /S/) presented in babble noise. They compared "clear" speech productions with "conversational" speech. The intelligibility was measured by determining the signal-to-noise ratio (SNR) at which listeners could achieve 75% accuracy. A key methodological aspect was the use of an adaptive procedure to efficiently estimate these thresholds.

Experiment 2 replicated Experiment 1 with listeners who had simulated hearing loss. This simulation involved a specific process to mimic the effects of sloping, recruiting hearing loss by altering the spectral envelope of the speech signal. This allowed the researchers to investigate how different auditory profiles interact with clear speech.

To understand why clear speech helps, the researchers performed correlation analyses. They measured various acoustic features of the fricatives (e.g., spectral peak locations, energy distribution, duration) in both clear and conversational speech for each speaker. They then correlated the differences in these acoustic features with the differences in intelligibility (the "clear speech benefit") across speakers.

Datasets

The study utilized a database of 8800 VCV ([a]-fricative-[a]) stimuli produced by 20 speakers (10 male, 10 female). These stimuli were part of a previous acoustic study on clear fricatives. The stimuli were presented in a background of 12-talker babble noise. No new datasets were introduced; rather, existing stimuli were used and manipulated for the experiments.

Key Insights and Ideas

  • Clear Speech is Generally Beneficial, But Not Universally: While clear speech improved intelligibility for both normal-hearing and simulated hearing-impaired listeners, the effect was not uniform across all fricative types or listener groups.
  • Different Acoustic Cues Matter for Different Listeners: For normal-hearing listeners, intelligibility gains in clear speech were linked to acoustic changes that shifted energy towards higher frequencies. However, for listeners with simulated hearing loss, these higher-frequency shifts were less beneficial, and in some cases, even detrimental. This suggests that the acoustic features that enhance intelligibility are audibility-dependent.
  • Sibilants vs. Non-Sibilants Behave Differently: Sibilant fricatives (/s/, /S/, /z/, /3/) showed a more consistent clear speech benefit than non-sibilant fricatives (/f/, /v/, /θ/, /ð/), particularly for normal-hearing listeners.
  • Simulated Hearing Loss Significantly Increases Difficulty: Listeners with simulated hearing loss required significantly higher SNRs to achieve the same level of intelligibility compared to normal-hearing listeners, highlighting the substantial impact of hearing impairment.

Key Results

  • Experiment 1 (Normal Hearing): Clear speech significantly lowered SNR thresholds across most fricative pairs (p<.0001). The "clear speech effect" was particularly strong for sibilant pairs. Acoustic correlates indicated that shifts in spectral energy to higher frequencies were associated with intelligibility gains.
  • Experiment 2 (Simulated Hearing Loss): Clear speech also provided a benefit (p<.01), but the effect was less pronounced and varied by fricative pair. Notably, for the /f/-/θ/ pair, clear speech actually resulted in higher (less intelligible) thresholds, an "anti-clear speech effect." The interaction between speaking style, listener group, and fricative pair was significant (p<.05), showing that the clear speech benefit for sibilants increased in the simulated loss condition, while it decreased for non-sibilants.
  • Group Comparison: Listeners with simulated hearing loss had considerably higher thresholds (4.47 dB) than normal-hearing listeners.

Limitations and Future Work

The authors acknowledge that the simulated hearing loss is a simplification and may not perfectly capture the complexities of actual hearing impairments, such as variations in temporal resolution or spectral sensitivity. They also note that the correlation analyses were less consistent for the simulated hearing loss group, and some results even suggested an inverse relationship between higher-frequency acoustic modifications and intelligibility for this group.

Future work could involve more controlled blocked-design experiments to better power analyses of specific fricatives, styles, and SNRs across speakers. The authors suggest that understanding these acoustic correlates can inform the design of intelligibility enhancement techniques, such as auditory training or prosthetic hearing devices, potentially by tailoring strategies to individual listener profiles and the specific acoustic characteristics of different speech sounds.

Summary updated

BibTeX
@inproceedings{maniwa2007acoustic,
  author = {Kazumi Maniwa and Allard Jongman and Travis Wade},
  title = {{Acoustic correlates of intelligibility enhancements in clearly produced fricatives}},
  booktitle = {Interspeech},
  year = {2007},
}
Figure