10.1. Introduction

Progress made in automatic speech recognition during the past decades has resulted in systems providing high performance levels in well-controlled operating conditions. Such systems are currently used in various domains: office automation (dictation machines), Internet applications, voice command, etc. [RAB 93]. The level of performance is strongly dependent on the complexity and the level of difficulty of the application under consideration. Thus, typical laboratory error rates in speaker-independent mode range from 0.3% (for digit sequences), 5% (for 20,000 word continuous speech recognition), up to 8% for spelled letter, and even above 50% for spontaneous telephone conversations!

In addition, the error rate tends to increase dramatically where there is a mismatch between the training conditions and the actual conditions of use (especially noise levels [MOK 98]). For example, experiments [SIO 95] have shown that, for a system trained with clean speech, the continuous speech recognition error rate varies from 3% with a very low SNR (signal-to-noise ratio) of 36 dB, up to 97% with a very noisy signal with a SNR = 0 dB (note: in these experiments, the noise is additive, white and Gaussian).

Current systems are therefore very sensitive to variability factors, even when they seem minor to the ear and sometimes even completely imperceptible. The current systems clearly lack robustness.

Variability factors in speech can be classified into three categories, depending on their sources, which include:

• the environment of the speaker: noise correlated to speech (reverberation, reflections) or additive (surrounding...