Automated Speech Audiometry: Can It Work Using Open-Source Pre-Trained Kaldi-NL Automatic Speech Recognition?
Gloria Araiza-Illan, Luke Meyer, Khiet P Truong and 1 others
PMID 38483979WHAT IT FOUND
For 30 normal-hearing native Dutch adults, an open-source speech recognizer scored the digits-in-noise test with average recognition errors 5.0%.
One person had 48% errors, and people with hearing loss were not tested.
Key findings
01The average word error rate was 5.0%, but one participant had 48% errors.
02Excluding the outlier, word error rates were 13.5% or lower, and six participants had 0% errors.
03In the simulation, inserting up to four triplets with decoding errors changed the DIN test score by 0.71 dB or less.
STILL TO COME
How it was doneWhat they foundWhat it means for SLPs
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
The test was run only in 30 self-reported normal-hearing native Dutch adults, so it does not show how the setup performs for people with hearing loss, children, or non-native speakers. One participant had a 48% word error rate, and the paper notes this person spoke quickly and quietly. The setup was tested without a human supervisor, and the effect of decoding errors on real clinical scores was modelled rather than measured against a human-scored DIN test. The room was quiet but not sound-treated, and the laptop speakers had lower levels in bands below 630 Hz. The authors state that audio quality, Dutch accent, and lack of speech context can reduce accuracy. Study 2 used only the six participants with 0% word error rate, so the simulation starts from the best cases.
Declared interests
The authors declared no potential conflicts of interest. Funding came from VICI grant 918-17-603 from the Netherlands Organization for Scientific Research and the Netherlands Organization for Health Research and Development, with further support from the Heinsius Houbolt Foundation and the W.J. Kolff Institute for Biomedical Engineering and Materials Science at University Medical Center Groningen.
The easy way to misread this
Do not assume automated scoring is ready for your patients. It was tested only in 30 normal-hearing native Dutch adults, and one participant had 48% recognition errors.