Using automatic speech recognition to improve hearing-aid fitting

Because people with age-related hearing loss (ARHL) generally experience difficulties in understanding speech, tests of speech identification are often used by audiologists and hearing-aid (HA) dispensers to evaluate the benefits of rehabilitation through hearing-aids.
However, these tests are fairly time-consuming, which can lead to an increase in fatigue (and thus potentially to a decrease in performance) for the older listeners. Moreover, the listeners’ speech-identification performances are likely to be influenced by their familiarity with the speech materials, which, ideally, should be refreshed for every test condition. These issues make it impossible to test all the HA settings that might yield optimal speech intelligibility for the listener.
Automatic speech recognition (ASR) systems could overcome these shortcomings, by providing fast and objective measures of speech intelligibility, provided that the perceptual consequences of ARHL can be accurately simulated by signal-processing algorithms.
In this presentation, we report on a series of proof-of-concept experiments that compared human speech-identification performances with the performance of an ASR system fed with speech signals simulating three of the perceptual consequences of ARHL (Nejime & Moore, 1997): elevation of hearing thresholds, loss of frequency selectivity, and loudness recruitment. For both young, normal-hearing listeners (Fontan et al., 2017), and older, hearing-impaired listeners (Fontan, Cretin-Maitenaz, & Füllgrabe, in revision), strong correlations between human and ASR performance were observed, indicating that trends in speech intelligibility can be predicted.
The system was later used in combination with a HA simulator (Moore, Füllgrabe & Stone, 2010) to find optimal HA amplification gains (in terms of predicted intelligibility) for 24 older listeners with ARHL. Participants’ aided speech-intelligibility scores and subjective judgements of speech pleasantness were found to be significantly higher when applying the ASR-based amplification gains than when applying a baseline fitting rule (Moore, Glasberg, & Stone, 2010).

References

Fontan, Cretin-Maitenaz, & Füllgrabe. (In revision). Predicting speech perception in older listeners with sensorineural hearing loss using automatic speech recognition. Trends in Hearing.
Fontan, Ferrané, Farinas, Pinquier, Magnen, Tardieu, Gaillard, Aumont, & Füllgrabe. (2017). Automatic speech recognition predicts speech intelligibility and comprehension for listeners with simulated age-related hearing loss. JSLHR, 60(9), 2394–2405.
Moore, Füllgrabe, & Stone. (2010) Effect of spatial separation, extended bandwidth, and compression speed on intelligibility in a competing-speech task. JASA, 128(1), 360–371.
Moore, Glasberg, & Stone. (2010). Development of a new method for deriving initial fittings for hearing aids with multi-channel compression: CAMEQ2-HF. IJA, 49(3), 216–227.
Nejime, & Moore. (1997). Simulation of the effect of threshold elevation and loudness recruitment combined with reduced frequency selectivity on the intelligibility of speech in noise. JASA, 102(1), 603–615.