Speech intelligibility is most accurately and representatively measured using auditory procedures, involving panels of human test subjects. If speech intelligibility has to be determined for a specific population (such as people suffering from hearing impairment), one simply recruits test subjects from this specific population. Unfortunately, auditory tests are cumbersome and expensive. For this reason, researchers, engineers and acoustics consultants often rely on instrumental procedures to predict speech intelligibility. Examples of such procedures are the Articulation Index (AI), Speech Intelligibility Index (SII), and the Speech Transmission Index (STI).
The SII and the STI are considered to represent the state of the art in intelligibility prediction. Although clear differences exist between these models (related to the models themselves and to their applications), both methods have many common features. The models are based on the observation that information carried in speech can be thought of as the sum of contributions by individual frequency bands. The models incorporate knowledge about peripheral auditory analysis and relative importance of individual frequency bands to the overall intelligibility to arrive at intelligibility predictions in form of a 0-1 index that is easy to interpret.
Standardized versions of the STI and SII are monaural models, in the sense that they are based on single-channel estimates. SII and STI were designed to predict intelligibility in diotic ('mono') listening conditions based on measurements with a single microphone. This means that binaural (dichotic) intelligibility benefits are disregarded. The benefit of listening to speech with two ears instead of one in conditions with background babble is known as the cocktail party effect. Scientific research for more than 50 years has yielded ample evidence for this cocktail party effect. Although the binaural benefit was demonstrated in many occasions, things are different for hearing impaired listeners. For one thing, differences in hearing loss between both ears may be considerable; binaural unmasking effects may be completely different from normal hearing listeners.
Preliminary listening tests with normal hearing and hearing impaired subjects have shown that binaural models give good predictions of speech intelligibility in (simulated) adverse acoustical environments. Future work will focus on fine-tuning of model parameters, implementation of (faster) algorithms, and more elaborate validation in real-world scenarios. Eventually, the algorithms will be adapted/extended to model intelligibility by vulnerable groups.
In an Auditory Virtual Environment (AVE) humans have auditory perceptions that do not correspond to their real environment but to a virtual one. An AVE generator can be used to demonstrate communication conditions of various difficulties in virtual daily life situations. As communication in adverse conditions is especially difficult for hearing impaired persons, an AVE could also be used for self-screening tests. Such situations occur when several persons are talking at the same time (e.g., cocktail party or cafeteria), in the presence of loud background noise (e.g., traffic in the street), or in environments with strong reverberation (like in big rooms, e.g., a train station).
One of our aims is to provide such self-screening tests over the Internet and to raise awareness on difficult hearing situations. If in such a self-screening listening test the answers deviate from the answers of a normal hearing reference group, a hearing impairment is indicated. In this case, the test person is advised to seek professional help at an audiologist.