Evaluation of the Lombard effect model on synthesizing Lombard speech in varying noise level environments with limited data

Lombard speech is intelligible speech produced by humans in noises. In this study, we focus on mimicking Lombard speech from natural neutral speech under backgrounds with varying noise levels to increase its intelligibility in these noises. Other approaches map corresponding speech features from the neutral speech to Lombard speech, which can only apply for an individual noise level, and cannot reveal feature tendencies. Instead, we implement a Lombard effect model to continuously estimate feature values with varying noise levels. The techniques, which are based on coarticulation, a source-filter model with MRTD and spectral-GMM, are used to easily modify features of the neutral speech to obtain their tendencies. Finally, these features are synthesized by STRAIGHT vocoder to obtain Lombard speech. The mimicking quality is evaluated in subjective listening experiments on similarity, naturalness, and intelligibility. The evaluation results show that the proposed method could convert neutral speech into Lombard speech in varying noise levels, which obtains comparable results with the state-of-the-art method.