Inventory based speech denoising with hidden Markov models

We are presenting a new speech waveform inventory based approach for the denoising of speech. The method combines an inventory style parametric description of speech signals with a statistical analysis of the underlying parameter space in clean and noisy conditions. Sufficient parameter statistics for successful denoising can be learned from around 40 minutes of (clean speech) training data. Shorter training sets are feasible, but may lead to quality reductions. Manual transcription of the training data is not required. The proposed procedure is intended for applications in which speaker enrollment and noise enrollment are feasible. Such applications include vehicular speaker-phone communication systems and jet pilot communication systems. The proposed method compares very favorably to commonly used waveform based denoising methods in both objective and subjective speech quality assessments.