Unsupervised Training of Neural Mask-Based Beamforming

Abstract

We present an unsupervised training approach for a neural network-based mask estimator in an acoustic beamforming application. The network is trained to maximize a likelihood criterion derived from a spatial mixture model of the observations. It is trained from scratch without requiring any parallel data consisting of degraded input and clean training targets. Thus, training can be carried out on real recordings of noisy speech rather than simulated ones. In contrast to previous work on unsupervised training of neural mask estimators, our approach avoids the need for a possibly pre-trained teacher model entirely. We demonstrate the effectiveness of our approach by speech recognition experiments on two different datasets: one mainly deteriorated by noise (CHiME 4) and one by reverberation (REVERB). The results show that the performance of the proposed system is on par with a supervised system using oracle target masks for training and with a system trained using a model-based teacher.

Integration of Neural Networks and Probabilistic Spatial Models for Acoustic Blind Source Separation

“…In contrast, the latter system is better tuned towards performance relying on a state of the art spatial clustering model and reporting competitive WERs. In [177] we presented a different take on unsupervised mask estimation: the likelihood under the assumption that the data follows a cACGMM is used as a maximization criterion to train a neural network which just provides the initialization to a single EM-step of the cACGMM parameter estimation process. That way, the network is encouraged to provide a mask as initialization which is close to an optimal initialization, thus leading to a higher likelihood.…”

Deep Bayesian Unsupervised Source Separation Based On A Complex Gaussian Mixture Model

“…Seetharaman et al [13] designed a loss function weighted by a confidence measure of the estimated references. Drude et al [15] also proposed a novel approach that directly trains a separation network from the cACGMM likelihood. They applied the method to noisy speech recordings and reported that the performance of automatic speech recognition was superior to that of their previous approach mentioned above.…”

“…Unsupervised training for neural source separation using multichannel mixture signals has recently gained a lot of attention [12][13][14][15]. One approach is to generate supervised data by using multichannel separation methods [12][13][14].…”

“…This approach suffers from the estimation errors of the multichannel methods mentioned above. To solve this problem, Drude et al [15] trained a separation network by directly optimizing the likelihood function of a cGMM. They reported that the performance of a conventional multichannel method was improved by initializing it with the network output.…”

A Style Transfer Approach to Source Separation

“…Consequently, supervised source separation using neural networks relies on the availability of paired mixtureclean data in the training set and cannot be used when such paired datasets are not available or expensive to collect. To relax these constraints, a few recent papers use other forms of information like the spatial separation between the sources in a multi-microphone setting, to train the networks for unsupervised source separation [7,8,9,10]. However, these constraints continue to impose restrictions on singlechannel source separation, where such secondary forms of information about the sources are not available.…”