3-1 Speech Intelligibility

Intelligibility model:

It takes into account room acoustics characteristics. To this end, the impulse response of the room is studied: its measurement, its role in estimating scores for intelligibility and for quality criteria. The work has allowed to propose a model to predict intelligibility scores based on the impulse response of the local.

The originality of the model is to consider separately the influence of the transducer and those of the room. Until then, all models used the calculation of early energy (useful) and late one (harmful) from an estimate or a measure of the overall impulse response (which is the convolution of the loudspeaker one with that of the microphone and that of the room at the measurement position).

Here, the model provides the contribution of speaker to intelligibility in terms of lack of bandwidth  and an energy ratio calculated from the overall impulse response and the impulse response of the deconvoluted room effect. This ratio contains the coupling between the diffuser and the radiation properties of the first acoustic reflections in the room. Simulation of changes in these parameters is in good agreement with the empirical results that are known to improve intelligibility scores with increasing the directivity of the sound emitter and increased sound absorption walls.

The model is validated by tests in highly reverberant room and in the presence of noise. The goal is then to introduce the action of (or) transducer (s) receptor (s) in the model. Until the models had no other purpose than to predict the scores of intelligibility in the room where was the voice source. Consequently, the impulse response estimates were always made with omnidirectional microphones, and then introduced into the models for comparison with the scores measured in situ with listeners.

Here, the speech messages are recorded by directional microphones like the acquisition of the pseudo-random binary sequences to calculate the impulse responses. The recordings are then distributed in another room for the measurement and modeling scores. The analyzes lead to two results.

– The intelligibility of recorded messages with directional microphones is better than that of the messages heard in the room at the same points. The sound (mono and stereo) therefore improves the understanding of speech.

– The model underestimates the scores for listening to recordings while maintaining their sense of variation with distance from the source of speech. The test allowed us to propose a method for simulating a sonogram. The echogram obtained with a cardioid can be calculated from the measurement of two measured echograms respectively with an omnidirectional microphone and a bidirectional microphone.

Applications: work on speech intelligibility were applied for the prediction of site of directivity characteristics of a speaker to estimate the intelligibility with the model developed by the laboratory (partner company Bouyer / Tyco). A method is proposed and validated for measures to overcome the anechoic consider leaving a forecast of intelligibility scores in situ with broadcasters in place. Meanwhile, we had to proceed to the choice of sound diffuser to achieve the best scores of intelligibility in a subway station (partner: Paris Transport Authority RATP). The output from the model that incorporates the features guidelines, frequency response of broadcasters and the impulse responses of the local, the choice has been made. The study has allowed comparison of the transducers not only in terms of electro but in terms of intelligibility for potential binding conditions of noise and reverberation.