Electrical Engineering And Systems Science

This paper introduces a novel neural network-based speech coding system that can process noisy speech effectively. The proposed source-aware neural audio coding (SANAC) system harmonizes a deep autoencoder-based source separation model and a neural coding system so that it can explicitly perform source separation and coding in the latent space. An added benefit of this system is that the codec can allocate a different amount of bits to the underlying sources so that the more important source sounds better in the decoded signal. We target a new use case where the user on the receiver side cares about the quality of the non-speech components in speech communication, while the speech source still carries the most crucial information. Both objective and subjective evaluation tests show that SANAC can recover the original noisy speech better than the baseline neural audio coding system, which is with no source-aware coding mechanism, and two conventional codecs.

Recent advancements in deep learning led to human-level performance in single-speaker speech synthesis. However, there are still limitations in terms of speech quality when generalizing those systems into multiple-speaker models especially for unseen speakers and unseen recording qualities. For instance, conventional neural vocoders are adjusted to the training speaker and have poor generalization capabilities to unseen speakers. In this work, we propose a variant of WaveRNN, referred to as speaker conditional WaveRNN (SC-WaveRNN). We target towards the development of an efficient universal vocoder even for unseen speakers and recording conditions. In contrast to standard WaveRNN, SC-WaveRNN exploits additional information given in the form of speaker embeddings. Using publicly-available data for training, SC-WaveRNN achieves significantly better performance over baseline WaveRNN on both subjective and objective metrics. In MOS, SC-WaveRNN achieves an improvement of about 23% for seen speaker and seen recording condition and up to 95% for unseen speaker and unseen condition. Finally, we extend our work by implementing a multi-speaker text-to-speech (TTS) synthesis similar to zero-shot speaker adaptation. In terms of performance, our system has been preferred over the baseline TTS system by 60% over 15.5% and by 60.9% over 32.6%, for seen and unseen speakers, respectively.

Speaker diarization is one of the actively researched topics in audio signal processing and machine learning. Utterance clustering is a critical part of a speaker diarization task. In this study, we aim to improve the performance of utterance clustering by processing multichannel (stereo) audio signals. We generated processed audio signals by combining left- and right-channel audio signals in a few different ways and then extracted embedded features (also called d-vectors) from those processed audio signals. We applied the Gaussian mixture model (GMM) for supervised utterance clustering. In the training phase, we used a parameter sharing GMM to train the model for each speaker. In the testing phase, we selected the speaker with the maximum likelihood as the detected speaker. Results of experiments with real audio recordings of multi-person discussion sessions showed that our proposed method that used multichannel audio signals achieved significantly better performance than a conventional method with mono audio signals.

In this study, we propose the global context guided channel and time-frequency transformations to model the long-range, non-local time-frequency dependencies and channel variances in speaker representations. We use the global context information to enhance important channels and recalibrate salient time-frequency locations by computing the similarity between the global context and local features. The proposed modules, together with a popular ResNet based model, are evaluated on the VoxCeleb1 dataset, which is a large scale speaker verification corpus collected in the wild. This lightweight block can be easily incorporated into a CNN model with little additional computational costs and effectively improves the speaker verification performance compared to the baseline ResNet-LDE model and the Squeeze&Excitation block by a large margin. Detailed ablation studies are also performed to analyze various factors that may impact the performance of the proposed modules. We find that by employing the proposed L2-tf-GTFC transformation block, the Equal Error Rate decreases from 4.56% to 3.07%, a relative 32.68% reduction, and a relative 27.28% improvement in terms of the DCF score. The results indicate that our proposed global context guided transformation modules can efficiently improve the learned speaker representations by achieving time-frequency and channel-wise feature recalibration.

Target speech extraction, which extracts the speech of a target speaker in a mixture given auxiliary speaker clues, has recently received increased interest. Various clues have been investigated such as pre-recorded enrollment utterances, direction information, or video of the target speaker. In this paper, we explore the use of speaker activity information as an auxiliary clue for single-channel neural network-based speech extraction. We propose a speaker activity driven speech extraction neural network (ADEnet) and show that it can achieve performance levels competitive with enrollment-based approaches, without the need for pre-recordings. We further demonstrate the potential of the proposed approach for processing meeting-like recordings, where the speaker activity is obtained from a diarization system. We show that this simple yet practical approach can successfully extract speakers after diarization, which results in improved ASR performance, especially in high overlapping conditions, with a relative word error rate reduction of up to 25%.

Speaker attribution is required in many real-world applications, such as meeting transcription, where speaker identity is assigned to each utterance according to speaker voice profiles. In this paper, we propose to solve the speaker attribution problem by using graph-based semi-supervised learning methods. A graph of speech segments is built for each session, on which segments from voice profiles are represented by labeled nodes while segments from test utterances are unlabeled nodes. The weight of edges between nodes is evaluated by the similarities between the pretrained speaker embeddings of speech segments. Speaker attribution then becomes a semi-supervised learning problem on graphs, on which two graph-based methods are applied: label propagation (LP) and graph neural networks (GNNs). The proposed approaches are able to utilize the structural information of the graph to improve speaker attribution performance. Experimental results on real meeting data show that the graph based approaches reduce speaker attribution error by up to 68% compared to a baseline speaker identification approach that processes each utterance independently.

Acoustic-to-articulatory inversion (AAI) methods estimate articulatory movements from the acoustic speech signal, which can be useful in several tasks such as speech recognition, synthesis, talking heads and language tutoring. Most earlier inversion studies are based on point-tracking articulatory techniques (e.g. EMA or XRMB). The advantage of rtMRI is that it provides dynamic information about the full midsagittal plane of the upper airway, with a high 'relative' spatial resolution. In this work, we estimated midsagittal rtMRI images of the vocal tract for speaker dependent AAI, using MGC-LSP spectral features as input. We applied FC-DNNs, CNNs and recurrent neural networks, and have shown that LSTMs are the most suitable for this task. As objective evaluation we measured normalized MSE, Structural Similarity Index (SSIM) and its complex wavelet version (CW-SSIM). The results indicate that the combination of FC-DNNs and LSTMs can achieve smooth generated MR images of the vocal tract, which are similar to the original MRI recordings (average CW-SSIM: 0.94).

Articulatory-to-acoustic (forward) mapping is a technique to predict speech using various articulatory acquisition techniques (e.g. ultrasound tongue imaging, lip video). Real-time MRI (rtMRI) of the vocal tract has not been used before for this purpose. The advantage of MRI is that it has a high `relative' spatial resolution: it can capture not only lingual, labial and jaw motion, but also the velum and the pharyngeal region, which is typically not possible with other techniques. In the current paper, we train various DNNs (fully connected, convolutional and recurrent neural networks) for articulatory-to-speech conversion, using rtMRI as input, in a speaker-specific way. We use two male and two female speakers of the USC-TIMIT articulatory database, each of them uttering 460 sentences. We evaluate the results with objective (Normalized MSE and MCD) and subjective measures (perceptual test) and show that CNN-LSTM networks are preferred which take multiple images as input, and achieve MCD scores between 2.8-4.5 dB. In the experiments, we find that the predictions of speaker `m1' are significantly weaker than other speakers. We show that this is caused by the fact that 74% of the recordings of speaker `m1' are out of sync.

Does speaking style variation affect humans' ability to distinguish individuals from their voices? How do humans compare with automatic systems designed to discriminate between voices? In this paper, we attempt to answer these questions by comparing human and machine speaker discrimination performance for read speech versus casual conversations. Thirty listeners were asked to perform a same versus different speaker task. Their performance was compared to a state-of-the-art x-vector/PLDA-based automatic speaker verification system. Results showed that both humans and machines performed better with style-matched stimuli, and human performance was better when listeners were native speakers of American English. Native listeners performed better than machines in the style-matched conditions (EERs of 6.96% versus 14.35% for read speech, and 15.12% versus 19.87%, for conversations), but for style-mismatched conditions, there was no significant difference between native listeners and machines. In all conditions, fusing human responses with machine results showed improvements compared to each alone, suggesting that humans and machines have different approaches to speaker discrimination tasks. Differences in the approaches were further confirmed by examining results for individual speakers which showed that the perception of distinct and confused speakers differed between human listeners and machines.

In this paper, we study a novel technique that exploits the interaction between speaker traits and linguistic content to improve both speaker verification and utterance verification performance. We implement an idea of speaker-utterance dual attention (SUDA) in a unified neural network. The dual attention refers to an attention mechanism for the two tasks of speaker and utterance verification. The proposed SUDA features an attention mask mechanism to learn the interaction between the speaker and utterance information streams. This helps to focus only on the required information for respective task by masking the irrelevant counterparts. The studies conducted on RSR2015 corpus confirm that the proposed SUDA outperforms the framework without attention mask as well as several competitive systems for both speaker and utterance verification.