Speech and Vision Lab

Speaking in the presence of noise changes the characteristics of the speech produced which is known as the Lombard effect. This effect is perceptually felt with an increase in intensity of speaking. These changes in the characteristics of speech production is to en- sure an intelligible communication in noisy environment. These changes also result in the performance degradation of speech systems like speaker recognition, speech recognition, etc. Human speech production mechanism is affected due to the Lombard effect, and is reflected mainly in the excitation source. Previous studies have focussed mainly on the changes in system level features. In our work, we examine the excitation source to study its changes due to Lombard effect. The excitation source features studied in this work are the instantaneous fundamental frequency F0 (i.e., pitch), the strength of excitation at the instants of significant excita- tion and a loudness measure reflecting the sharpness of the impulse-like excitation around epochs. Another feature, called the normalized energy is used to study the articulation variability of speech due to Lombard effect using the high variations of energy within an utterance. Analysis is performed using the distributions of the features which are seen to distinguish between normal speech and Lombard effect speech, and also to explain the speaker-specific nature of Lombard effect. The extent of Lombard effect on speech depends on the nature and intensity of the noise that causes the Lombard effect. Charac- teristics of speech produced without a feedback and difference between loud speech and Lombard effect speech are also studied using the excitation source features. Duration is used to study the phonetic changes due to Lombard effect. Intelligibility due to Lombard effect at a sentence level is studied and the mechanism of the Lombard effect is described based on the analysis performed. This analysis is extended into an application where the imposters during speaker verification can be detected with the theory that speech pro- duced under noise is different from that spoken under normal conditions. Further, the performance of text-dependent speaker verification system due to Lombard effect is also described.