Front-end processing for the CHiME-5 dinner party scenario

Abstract

This contribution presents a speech enhancement system for the CHiME-5 Dinner Party Scenario. The front-end employs multi-channel linear time-variant filtering and achieves its gains without the use of a neural network. We present an adaptation of blind source separation techniques to the CHiME-5 database which we call Guided Source Separation (GSS). Using the baseline acoustic and language model, the combination of Weighted Prediction Error based dereverberation, guided source separation, and beamforming reduces the WER by 10.54 % (relative) for the single array track and by 21.12 % (relative) on the multiple array track.

An Investigation into the Effectiveness of Enhancement in ASR Training and Test for Chime-5 Dinner Party Transcription

“…GSS enhancement is a blind source separation technique originally proposed in [14] 1 to alleviate the speaker overlap problem in CHiME-5. Given a mixture of reverberated overlapped speech, GSS aims to separate the sources using a pure signal processing approach.…”

“…Those masks are also given by the mentioned class affiliations. For the single array (CHiME-5) track, simulations have shown that multiplying the beamformer output with the target speaker mask improves the performance on the U data, but the same approach degrades the performance in the multiple array track [14]. This is because the spatial selectivity of a single array is very limited in CHiME-5: the speakers' signals arrive at the array, which is mounted on the wall at some distance, at very similar impinging angles, rendering single array beamforming rather ineffective.…”

“…Temporal context also plays an important role in the EM initialization. Simulations have shown that a large context of 15 seconds left and right of the considered segment improves the mixture model estimation performance significantly for CHiME-5 [14]. However, having such a large temporal context may become problematic when the speakers are moving, because the estimated spatial covariance matrix can become outdated due to the movement [13].…”

“…In Table 1, the strength of the enhancement increases from top to bottom, i.e., GSS6 signals are much cleaner than the unprocessed ones. Single/Multi None BeamformIt [22] Single BFIt WPE + BeamformIt [10] Single WPE+BFIt WPE + GSS1 + BF w/o Context [14] Single GSS1 WPE + GSS6 + BF w/o Context [14] Multi GSS6 The standard CHiME-5 recipes were used to: (i) train GMM-HMM alignment models, (ii) clean up the training data, and (iii) augment the training data using three-fold speed perturbation. The acoustic feature vector consisted of 40-dimensional MFCCs appended with 100-dimensional i-vectors.…”

“…Recently, Guided Source Separation (GSS) enhancement on the test data was shown to significantly improve the performance of an acoustic model, which had been trained with a large amount of unprocessed and simulated noisy data [13]. GSS is a spatial mixture model based blind source separation approach which exploits the annotation given in the CHiME-5 database for initialization and, in this way, avoids the frequency permutation problem [14].…”