Multi-style training of HMMS with stereo data for reverberation-robust speech recognition

Abstract

A novel training algorithm using data pairs of clean and reverberant feature vectors for estimating robust Hidden Markov Models (HMMs), introduced in [1] for matched training, is employed in this paper for multi-style training. The multi-style HMMs are derived from well-trained cleanspeech HMMs by aligning the clean data to the clean-speech HMM and using the resulting state-frame alignment to estimate the Gaussian mixture densities from the reverberant data of several different rooms. Thus, the temporal alignment is fixed for all reverberation conditions contained in the multi-style training set so that the model mismatch between the different rooms is reduced. Therefore, this training approach is particularly suitable for multi-style training. Multi-style HMMs trained by the proposed approach and adapted to the current room condition using maximum likelihood linear regression significantly outperform the corresponding adapted multi-style HMMs trained by the conventional Baum-Welch algorithm. In strongly reverberant rooms, the proposed adapted multi-style HMMs even outperform Baum-Welch HMMs trained on matched data.

Single-Channel Speech Enhancement Techniques for Distant Speech Recognition

“…The increased recognition performance of ICEWIND over conventional Baum-Welch method is shown in Ref. [23].…”

“…In this case, only the RIR of the target room needs to be measured or estimated, and then a set of reverberant training data can be generated by convolving clean speech signals with this RIR. This concept is used in the information combining estimation with non-reverberant data (ICEWIND) Brought to you by | New York University Bobst Library Technical Services Authenticated Download Date | 6/21/15 2:18 PM method [23]. ICEWIND uses ordered pairs [s(k), x(k)] of clean and reverberant feature vector sequences as training data to determine the parameters λ x of a reverberant HMM.…”

Efficient training of acoustic models for reverberation-robust medium-vocabulary automatic speech recognition

“…This very accurate temporal alignment is then used to estimate the parameters of the emission densities of the reverberant HMM by applying the Expectation Maximization (EM) algorithm to the reverberant data. It has been shown in [5,7] that this approach significantly reduces the computational complexity of the HMM training and improves the recognition rates at the same time, both for room-specific and multi-style training of word-level HMMs for a connected-digit recognition task.…”

“…Promising approaches to reduce this mismatch are to train the acoustic models on corrupted data [1,2] and multi-style training [3][4][5][6]. In this paper, we follow these approaches using synthetically reverberated training data.…”

A Bayesian view on acoustic model-based techniques for robust speech recognition

“…For clarity, we entirely focus in this article on the compensation rules while ignoring the parameter estimation step. We also disregard approaches that apply a modified training method to conventional HMMs without exhibiting a distinct compensation step, as it is characteristic for, e.g., discriminative [28], multi-condition [29], or reverberant training [30].…”