Speaker embedding is an important front-end module to explore discriminative speaker features for many speech applications where speaker information is needed. Current SOTA backbone networks for speaker embedding are designed to aggregate multi-scale features from an utterance with multi-branch network architectures for speaker representation. However, naively adding many branches of multi-scale features with the simple fully convolutional operation could not efficiently improve the performance due to the rapid increase of model parameters and computational complexity. Therefore, in the most current state-of-the-art network architectures, only a few branches corresponding to a limited number of temporal scales could be designed for speaker embeddings. To address this problem, in this paper, we propose an effective temporal multi-scale (TMS) model where multi-scale branches could be efficiently designed in a speaker embedding network almost without increasing computational costs. The new model is based on the conventional TDNN, where the network architecture is smartly separated into two modeling operators: a channel-modeling operator and a temporal multi-branch modeling operator. Adding temporal multi-scale in the temporal multi-branch operator needs only a little bit increase of the number of parameters, and thus save more computational budget for adding more branches with large temporal scales. Moreover, in the inference stage, we further developed a systemic re-parameterization method to convert the TMS-based model into a single-path-based topology in order to increase inference speed. We investigated the performance of the new TMS method for automatic speaker verification (ASV) on in-domain and out-of-domain conditions. Results show that the TMS-based model obtained a significant increase in the performance over the SOTA ASV models, meanwhile, had a faster inference speed.
In acoustic signal processing, the target signals usually carry semantic information, which is encoded in a hierarchal structure of short and long-term contexts. However, the background noise distorts these structures in a nonuniform way. The existing deep acoustic signal enhancement (ASE) architectures ignore this kind of local and global effect. To address this problem, we propose to integrate a novel temporal attentive-pooling (TAP) mechanism into a conventional convolutional recurrent neural network, termed as TAP-CRNN. The proposed approach considers both global and local attention for ASE tasks. Specifically, we first utilize a convolutional layer to extract local information of the acoustic signals and then a recurrent neural network (RNN) architecture is used to characterize temporal contextual information. Second, we exploit a novelattention mechanism to contextually process salient regions of the noisy signals. The proposed ASE system is evaluated using a benchmark infant cry dataset and compared with several well-known methods. It is shown that the TAPCRNN can more effectively reduce noise components from infant cry signals in unseen background noises at challenging signal-to-noise levels.
This paper presents a novel discriminator-constrained optimal transport network (DOTN) that performs unsupervised domain adaptation for speech enhancement (SE), which is an essential regression task in speech processing. The DOTN aims to estimate clean references of noisy speech in a target domain, by exploiting the knowledge available from the source domain. The domain shift between training and testing data has been reported to be an obstacle to learning problems in diverse fields. Although rich literature exists on unsupervised domain adaptation for classification, the methods proposed, especially in regressions, remain scarce and often depend on additional information regarding the input data. The proposed DOTN approach tactically fuses the optimal transport (OT) theory from mathematical analysis with generative adversarial frameworks, to help evaluate continuous labels in the target domain. The experimental results on two SE tasks demonstrate that by extending the classical OT formulation, our proposed DOTN outperforms previous adversarial domain adaptation frameworks in a purely unsupervised manner.
Automatic speaker verification (ASV) systems, which determine whether two speeches are from the same speaker, mainly focus on verification accuracy while ignoring inference speed. However, in real applications, both inference speed and verification accuracy are essential. This study proposes cross-sequential re-parameterization (CS-Rep), a novel topology re-parameterization strategy for multi-type networks, to increase the inference speed and verification accuracy of models. CS-Rep solves the problem that existing re-parameterization methods are unsuitable for typical ASV backbones. When a model applies CS-Rep, the training-period network utilizes a multi-branch topology to capture speaker information, whereas the inference-period model converts to a time-delay neural network (TDNN)-like plain backbone with stacked TDNN layers to achieve the fast inference speed. Based on CS-Rep, an improved TDNN with friendly test and deployment called Rep-TDNN is proposed. Compared with the state-of-the-art model ECAPA-TDNN, which is highly recognized in the industry, Rep-TDNN increases the actual inference speed by about 50% and reduces the EER by 10%. The code will be released.
The discrepancy between the cost function used for training a speech enhancement model and human auditory perception usually makes the quality of enhanced speech unsatisfactory. Objective evaluation metrics which consider human perception can hence serve as a bridge to reduce the gap. Our previously proposed MetricGAN was designed to optimize objective metrics by connecting the metric with a discriminator. Because only the scores of the target evaluation functions are needed during training, the metrics can even be non-differentiable. In this study, we propose a MetricGAN+ in which three training techniques incorporating domain-knowledge of speech processing are proposed. With these techniques, experimental results on the VoiceBank-DEMAND dataset show that MetricGAN+ can increase PESQ score by 0.3 compared to the previous MetricGAN and achieve state-of-the-art results (PESQ score = 3.15).
Generative probability models are widely used for speaker verification (SV). However, the generative models are lack of discriminative feature selection ability. As a hypothesis test, the SV can be regarded as a binary classification task which can be designed as a Siamese neural network (SiamNN) with discriminative training. However, in most of the discriminative training for SiamNN, only the distribution of pair-wised sample distances is considered, and the additional discriminative information in joint distribution of samples is ignored. In this paper, we propose a novel SiamNN with consideration of the joint distribution of samples. The joint distribution of samples is first formulated based on a joint Bayesian (JB) based generative model, then a SiamNN is designed with dense layers to approximate the factorized affine transforms as used in the JB model. By initializing the SiamNN with the learned model parameters of the JB model, we further train the model parameters with the pair-wised samples as a binary discrimination task for SV. We carried out SV experiments on data corpus of speakers in the wild (SITW) and VoxCeleb. Experimental results showed that our proposed model improved the performance with a large margin compared with state of the art models for SV.
Synthesized speech from articulatory movements can have real-world use for patients with vocal cord disorders, situations requiring silent speech, or in high-noise environments. In this work, we present EMA2S, an end-to-end multimodal articulatory-to-speech system that directly converts articulatory movements to speech signals. We use a neural-network-based vocoder combined with multimodal joint-training, incorporating spectrogram, mel-spectrogram, and deep features. The experimental results confirm that the multimodal approach of EMA2S outperforms the baseline system in terms of both objective evaluation and subjective evaluation metrics. Moreover, results demonstrate that joint mel-spectrogram and deep feature loss training can effectively improve system performance.
The task for speaker verification (SV) is to decide an utterance is spoken by a target or imposter speaker. In most SV studies, a log-likelihood ratio (L_LLR) score is estimated based on a generative probability model on speaker features, and compared with a threshold for decision making. However, the generative model usually focuses on feature distributions and does not have the discriminative feature selection ability, which is easy to be distracted by nuisance features. The SV, as a hypothesis test, could be formulated as a binary classification task where a neural network (NN) based discriminative learning could be applied. Through discriminative learning, the nuisance features could be removed with the help of label supervision. However, the discriminative learning pays more attention to classification boundaries which is prone to overfitting to training data and yielding poor generalization on testing data. In this paper, we propose a hybrid learning framework, i.e., integrating a joint Bayesian (JB) generative model into a neural discriminative learning framework for SV. A Siamese NN is built with dense layers to approximate the mapping functions used in the SV pipeline with the JB model, and the L-LLR score estimated based on the JB model is connected to the distance metric in a pair-wised discriminative learning. By initializing the Siamese NN with the parameters learned from the JB model, we further train the model parameters with the pair-wised samples as a binary discrimination task. Moreover, direct evaluation metric in SV, i.e., minimum empirical Bayes risk, is designed and integrated as an objective function in the discriminative learning. We carried out SV experiments on speakers in the wild (SITW) and Voxceleb corpora. Experimental results showed that our proposed model improved the performance with a large margin compared with state-of-the-art models for SV.
Due to the mismatch of statistical distributions of acoustic speech between training and testing sets, the performance of spoken language identification (SLID) could be drastically degraded. In this paper, we propose an unsupervised neural adaptation model to deal with the distribution mismatch problem for SLID. In our model, we explicitly formulate the adaptation as to reduce the distribution discrepancy on both feature and classifier for training and testing data sets. Moreover, inspired by the strong power of the optimal transport (OT) to measure distribution discrepancy, a Wasserstein distance metric is designed in the adaptation loss. By minimizing the classification loss on the training data set with the adaptation loss on both training and testing data sets, the statistical distribution difference between training and testing domains is reduced. We carried out SLID experiments on the oriental language recognition (OLR) challenge data corpus where the training and testing data sets were collected from different conditions. Our results showed that significant improvements were achieved on the cross domain test tasks.
Speech enhancement (SE) aims to improve speech quality and intelligibility, which are both related to a smooth transition in speech segments that may carry linguistic information, e.g. phones and syllables. In this study, we took phonetic characteristics into account in the SE training process. Hence, we designed a phone-fortified perceptual (PFP) loss, and the training of our SE model was guided by PFP loss. In PFP loss, phonetic characteristics are extracted by wav2vec, an unsupervised learning model based on the contrastive predictive coding (CPC) criterion. Different from previous deep-feature-based approaches, the proposed approach explicitly uses the phonetic information in the deep feature extraction process to guide the SE model training. To test the proposed approach, we first confirmed that the wav2vec representations carried clear phonetic information using a t-distributed stochastic neighbor embedding (t-SNE) analysis. Next, we observed that the proposed PFP loss was more strongly correlated with the perceptual evaluation metrics than point-wise and signal-level losses, thus achieving higher scores for standardized quality and intelligibility evaluation metrics in the Voice Bank-DEMAND dataset.