lars_round
Lars Bramsløw

Research Engineer, PhD, Project Leader

> Learn more

mail@eriksholm.com

Existing speech tests are not optimal for assessing the effects of different hearing-aid fittings on speech understanding in multi-talker situations. Eriksholm developed new Danish speech material to fill this gap

Designing new speech material

Multi-talker listening situations are known to be particularly problematic for hearing-impaired (HI) listeners. Consequently, competing-speech tests are needed for evaluating the effects of different hearing-aid fittings on speech recognition in such situations.

A few years ago, we developed a multi-talker speech corpus for this purpose, based on the closed-set Dantale-II speech material1. However, closed-set materials typically result in fairly low (negative) speech reception thresholds (SRTs). Together with the Centre for Applied Hearing Research at the Technical University of Denmark, we therefore developed an open-set, multi-talker speech corpus for measuring higher (and therefore arguably more realistic) SRTs.

The new material is inspired by the English TVM corpus,2 with a sentence template similar to “Name thought about the noun and the noun yesterday”. The name is used for cuing the target sentence, and the nouns are used for scoring purposes. Based on this template, we created about 800 unique sentences, which were deviced and then recorded with the help of three professional female talkers in a professional recording studio.

Test list generation and validation

To be able to generate homogeneous test lists, we measured the intelligibility of the recorded sentences with 16 normal-hearing (NH) listeners. For each talker, we then selected the 200 sentences with the most similar intelligibilities and distributed them into 10 test lists of 20 sentences.

  • Designing new speech material fig 1

To validate these lists, another listening test with 16 NH listeners was conducted. For a speech target presented from the front, and two speech maskers presented simultaneously from ±50o, we obtained a mean SRT of approximately 1 dB – a value that is much higher than results typically obtained in other speech-on-speech tests. We also observed high test-retest reliability.

To complete the validation, reference data was collected on a group of hearing-impaired (HI) listeners for whom audibility had been ensured. For the same spatial test setup as that used for the NH listeners, we obtained a mean SRT of ca. 6 dB and a within-subject SD of ca. 1,4 dB.


Further reading

1 Behrens T, Neher T, Johannesson RB (2008). Evaluation of speech corpus for assessment of spatial release from masking. In: Dau T, Buchholz JM, Harte J. Christiansen TU (Eds.) Auditory Signal Processing in Hearing-Impaired Listeners. Proceedings of the First International Symposium on Audiological and Auditory Research, ISAAR, p. 449-457.

2 Helfer KS, Freyman RL (2009). Lexical and indexical cues in masking by competing speech. Journal of the Acoustical Society of America, 125, p. 447-456.

Learn more

  • learnmore-poster

    Poster

    New danish speech-on-speech material
  • studypartners_transparent

    Study Partners

    Jens Bo Nielsen and Torsten Dau from the Centre for Applied Hearing Research, DTU