Online Noise and Lombard Effect Compensation for In-Vehicle Automatic Speech Recognition

Presence of background noise in speech impacts the performance of automatic speech recognition (ASR). Adverse noisy environments are also known to induce so-called Lombard effect (LE), where speakers adjust their speech production in order to preserve intelligible communication. LE leads to further ASR degradation, often stronger than the one due to noise. Recently, a set of techniques reducing the impact of noise and LE have been introduced. In this paper, these algorithms are incorporated in a novel ASR setup and evaluated on neutral and Lombard speech corrupted by noise samples from a newly acquired car noise database. It is shown that the proposed scheme provides considerable performance improvement over the baseline and state-of-the-art approaches for all considered car environments and noise levels, reaching 4-17% absolute word error rate (WER) reduction.