State-of-the-art neural network language models (NNLMs) represented by long short term memory recurrent neural networks (LSTM-RNNs) and Transformers are becoming highly complex. They are prone to overfitting and poor generalization when given limited training data. To this end, an overarching full Bayesian learning framework encompassing three methods is proposed in this paper to account for the underlying uncertainty in LSTM-RNN and Transformer LMs. The uncertainty over their model parameters, choice of neural activations and hidden output representations are modeled using Bayesian, Gaussian Process and variational LSTM-RNN or Transformer LMs respectively. Efficient inference approaches were used to automatically select the optimal network internal components to be Bayesian learned using neural architecture search. A minimal number of Monte Carlo parameter samples as low as one was also used. These allow the computational costs incurred in Bayesian NNLM training and evaluation to be minimized. Experiments are conducted on two tasks: AMI meeting transcription and Oxford-BBC LipReading Sentences 2 (LRS2) overlapped speech recognition using state-of-the-art LF-MMI trained factored TDNN systems featuring data augmentation, speaker adaptation and audio-visual multi-channel beamforming for overlapped speech. Consistent performance improvements over the baseline LSTM-RNN and Transformer LMs with point estimated model parameters and drop-out regularization were obtained across both tasks in terms of perplexity and word error rate (WER). In particular, on the LRS2 data, statistically significant WER reductions up to 1.3% and 1.2% absolute (12.1% and 11.3% relative) were obtained over the baseline LSTM-RNN and Transformer LMs respectively after model combination between Bayesian NNLMs and their respective baselines.
Audio source separation is often achieved by estimating the magnitude spectrogram of each source, and then applying a phase recovery (or spectrogram inversion) algorithm to retrieve time-domain signals. Typically, spectrogram inversion is treated as an optimization problem involving one or several terms in order to promote estimates that comply with a consistency property, a mixing constraint, and/or a target magnitude objective. Nonetheless, it is still unclear which set of constraints and problem formulation is the most appropriate in practice. In this paper, we design a general framework for deriving spectrogram inversion algorithm, which is based on formulating optimization problems by combining these objectives either as soft penalties or hard constraints. We solve these by means of algorithms that perform alternating projections on the subsets corresponding to each objective/constraint. Our framework encompasses existing techniques from the literature as well as novel algorithms. We investigate the potential of these approaches for a speech enhancement task. In particular, one of our novel algorithms outperforms other approaches in a realistic setting where the magnitudes are estimated beforehand using a neural network.
Speaker verification (SV) suffers from unsatisfactory performance in far-field scenarios due to environmental noise andthe adverse impact of room reverberation. This work presents a benchmark of multichannel speech enhancement for far-fieldspeaker verification. One approach is a deep neural network-based, and the other is a combination of deep neural network andsignal processing. We integrated a DNN architecture with signal processing techniques to carry out various experiments. Ourapproach is compared to the existing state-of-the-art approaches. We examine the importance of enrollment in pre-processing,which has been largely overlooked in previous studies. Experimental evaluation shows that pre-processing can improve the SVperformance as long as the enrollment files are processed similarly to the test data and that test and enrollment occur within similarSNR ranges. Considerable improvement is obtained on the generated and all the noise conditions of the VOiCES dataset.
Research about brain activities involving spoken word production is considerably underdeveloped because of the undiscovered characteristics of speech artifacts, which contaminate electroencephalogram (EEG) signals and prevent the inspection of the underlying cognitive processes. To fuel further EEG research with speech production, a method using three-mode tensor decomposition (time x space x frequency) is proposed to perform speech artifact removal. Tensor decomposition enables simultaneous inspection of multiple modes, which suits the multi-way nature of EEG data. In a picture-naming task, we collected raw data with speech artifacts by placing two electrodes near the mouth to record lip EMG. Based on our evaluation, which calculated the correlation values between grand-averaged speech artifacts and the lip EMG, tensor decomposition outperformed the former methods that were based on independent component analysis (ICA) and blind source separation (BSS), both in detecting speech artifact (0.985) and producing clean data (0.101). Our proposed method correctly preserved the components unrelated to speech, which was validated by computing the correlation value between the grand-averaged raw data without EOG and cleaned data before the speech onset (0.92-0.94).
This paper presents our latest investigations on improving automatic speech recognition for noisy speech via speech enhancement. We propose a novel method named Multi-discriminators CycleGAN to reduce noise of input speech and therefore improve the automatic speech recognition performance. Our proposed method leverages the CycleGAN framework for speech enhancement without any parallel data and improve it by introducing multiple discriminators that check different frequency areas. Furthermore, we show that training multiple generators on homogeneous subset of the training data is better than training one generator on all the training data. We evaluate our method on CHiME-3 data set and observe up to 10.03% relatively WER improvement on the development set and up to 14.09% on the evaluation set.
Despite significant advances in deep models for music generation, the use of these techniques remains restricted to expert users. Before being democratized among musicians, generative models must first provide expressive control over the generation, as this conditions the integration of deep generative models in creative workflows. In this paper, we tackle this issue by introducing a deep generative audio model providing expressive and continuous descriptor-based control, while remaining lightweight enough to be embedded in a hardware synthesizer. We enforce the controllability of real-time generation by explicitly removing salient musical features in the latent space using an adversarial confusion criterion. User-specified features are then reintroduced as additional conditioning information, allowing for continuous control of the generation, akin to a synthesizer knob. We assess the performance of our method on a wide variety of sounds including instrumental, percussive and speech recordings while providing both timbre and attributes transfer, allowing new ways of generating sounds.
We present a textless speech-to-speech translation (S2ST) system that can translate speech from one language into another language and can be built without the need of any text data. Different from existing work in the literature, we tackle the challenge in modeling multi-speaker target speech and train the systems with real-world S2ST data. The key to our approach is a self-supervised unit-based speech normalization technique, which finetunes a pre-trained speech encoder with paired audios from multiple speakers and a single reference speaker to reduce the variations due to accents, while preserving the lexical content. With only 10 minutes of paired data for speech normalization, we obtain on average 3.2 BLEU gain when training the S2ST model on the \vp~S2ST dataset, compared to a baseline trained on un-normalized speech target. We also incorporate automatically mined S2ST data and show an additional 2.0 BLEU gain. To our knowledge, we are the first to establish a textless S2ST technique that can be trained with real-world data and works for multiple language pairs.
Recent advances in End-to-End (E2E) Spoken Language Understanding (SLU) have been primarily due to effective pretraining of speech representations. One such pretraining paradigm is the distillation of semantic knowledge from state-of-the-art text-based models like BERT to speech encoder neural networks. This work is a step towards doing the same in a much more efficient and fine-grained manner where we align speech embeddings and BERT embeddings on a token-by-token basis. We introduce a simple yet novel technique that uses a cross-modal attention mechanism to extract token-level contextual embeddings from a speech encoder such that these can be directly compared and aligned with BERT based contextual embeddings. This alignment is performed using a novel tokenwise contrastive loss. Fine-tuning such a pretrained model to perform intent recognition using speech directly yields state-of-the-art performance on two widely used SLU datasets. Our model improves further when fine-tuned with additional regularization using SpecAugment especially when speech is noisy, giving an absolute improvement as high as 8% over previous results.
Many studies have examined the shortcomings of word error rate (WER) as an evaluation metric for automatic speech recognition (ASR) systems, particularly when used for spoken language understanding tasks such as intent recognition and dialogue systems. In this paper, we propose Hybrid-SD ($\text{H}_{\text{SD}}$), a new hybrid evaluation metric for ASR systems that takes into account both semantic correctness and error rate. To generate sentence dissimilarity scores (SD), we built a fast and lightweight SNanoBERT model using distillation techniques. Our experiments show that the SNanoBERT model is 25.9x smaller and 38.8x faster than SRoBERTa while achieving comparable results on well-known benchmarks. Hence, making it suitable for deploying with ASR models on edge devices. We also show that $\text{H}_{\text{SD}}$ correlates more strongly with downstream tasks such as intent recognition and named-entity recognition (NER).
We present ClearBuds, the first hardware and software system that utilizes a neural network to enhance speech streamed from two wireless earbuds. Real-time speech enhancement for wireless earbuds requires high-quality sound separation and background cancellation, operating in real-time and on a mobile phone. Clear-Buds bridges state-of-the-art deep learning for blind audio source separation and in-ear mobile systems by making two key technical contributions: 1) a new wireless earbud design capable of operating as a synchronized, binaural microphone array, and 2) a lightweight dual-channel speech enhancement neural network that runs on a mobile device. Our neural network has a novel cascaded architecture that combines a time-domain conventional neural network with a spectrogram-based frequency masking neural network to reduce the artifacts in the audio output. Results show that our wireless earbuds achieve a synchronization error less than 64 microseconds and our network has a runtime of 21.4 milliseconds on an accompanying mobile phone. In-the-wild evaluation with eight users in previously unseen indoor and outdoor multipath scenarios demonstrates that our neural network generalizes to learn both spatial and acoustic cues to perform noise suppression and background speech removal. In a user-study with 37 participants who spent over 15.4 hours rating 1041 audio samples collected in-the-wild, our system achieves improved mean opinion score and background noise suppression. Project page with demos: https://clearbuds.cs.washington.edu