Self-Supervised Learning (SSL) based models of speech have shown remarkable performance on a range of downstream tasks. These state-of-the-art models have remained blackboxes, but many recent studies have begun "probing" models like HuBERT, to correlate their internal representations to different aspects of speech. In this paper, we show "inference of articulatory kinematics" as fundamental property of SSL models, i.e., the ability of these models to transform acoustics into the causal articulatory dynamics underlying the speech signal. We also show that this abstraction is largely overlapping across the language of the data used to train the model, with preference to the language with similar phonological system. Furthermore, we show that with simple affine transformations, Acoustic-to-Articulatory inversion (AAI) is transferrable across speakers, even across genders, languages, and dialects, showing the generalizability of this property. Together, these results shed new light on the internals of SSL models that are critical to their superior performance, and open up new avenues into language-agnostic universal models for speech engineering, that are interpretable and grounded in speech science.
Speech-to-face generation is an intriguing area of research that focuses on generating realistic facial images based on a speaker's audio speech. However, state-of-the-art methods employing GAN-based architectures lack stability and cannot generate realistic face images. To fill this gap, we propose a novel speech-to-face generation framework, which leverages a Speech-Conditioned Latent Diffusion Model, called SCLDM. To the best of our knowledge, this is the first work to harness the exceptional modeling capabilities of diffusion models for speech-to-face generation. Preserving the shared identity information between speech and face is crucial in generating realistic results. Therefore, we employ contrastive pre-training for both the speech encoder and the face encoder. This pre-training strategy facilitates effective alignment between the attributes of speech, such as age and gender, and the corresponding facial characteristics in the face images. Furthermore, we tackle the challenge posed by excessive diversity in the synthesis process caused by the diffusion model. To overcome this challenge, we introduce the concept of residuals by integrating a statistical face prior to the diffusion process. This addition helps to eliminate the shared component across the faces and enhances the subtle variations captured by the speech condition. Extensive quantitative, qualitative, and user study experiments demonstrate that our method can produce more realistic face images while preserving the identity of the speaker better than state-of-the-art methods. Highlighting the notable enhancements, our method demonstrates significant gains in all metrics on the AVSpeech dataset and Voxceleb dataset, particularly noteworthy are the improvements of 32.17 and 32.72 on the cosine distance metric for the two datasets, respectively.
The computational cost of transformer models makes them inefficient in low-latency or low-power applications. While techniques such as quantization or linear attention can reduce the computational load, they may incur a reduction in accuracy. In addition, globally reducing the cost for all inputs may be sub-optimal. We observe that for each layer, the full width of the layer may be needed only for a small subset of tokens inside a batch and that the "effective" width needed to process a token can vary from layer to layer. Motivated by this observation, we introduce the Adaptive Computation Module (ACM), a generic module that dynamically adapts its computational load to match the estimated difficulty of the input on a per-token basis. An ACM consists of a sequence of learners that progressively refine the output of their preceding counterparts. An additional gating mechanism determines the optimal number of learners to execute for each token. We also describe a distillation technique to replace any pre-trained model with an "ACMized" variant. The distillation phase is designed to be highly parallelizable across layers while being simple to plug-and-play into existing networks. Our evaluation of transformer models in computer vision and speech recognition demonstrates that substituting layers with ACMs significantly reduces inference costs without degrading the downstream accuracy for a wide interval of user-defined budgets.
In a spoken dialogue system, an NLU model is preceded by a speech recognition system that can deteriorate the performance of natural language understanding. This paper proposes a method for investigating the impact of speech recognition errors on the performance of natural language understanding models. The proposed method combines the back transcription procedure with a fine-grained technique for categorizing the errors that affect the performance of NLU models. The method relies on the usage of synthesized speech for NLU evaluation. We show that the use of synthesized speech in place of audio recording does not change the outcomes of the presented technique in a significant way.
In recent years, there have been studies to further improve the end-to-end neural speaker diarization (EEND) systems. This letter proposes the EEND-DEMUX model, a novel framework utilizing demultiplexed speaker embeddings. In this work, we focus on disentangling speaker-relevant information in the latent space and then transform each separated latent variable into its corresponding speech activity. EEND-DEMUX can directly obtain separated speaker embeddings through the demultiplexing operation in the inference phase without an external speaker diarization system, an embedding extractor, or a heuristic decoding technique. Furthermore, we employ a multi-head cross-attention mechanism to capture the correlation between mixture and separated speaker embeddings effectively. We formulate three loss functions based on matching, orthogonality, and sparsity constraints to learn robust demultiplexed speaker embeddings. The experimental results on the LibriMix dataset show consistently improved performance in both a fixed and flexible number of speakers scenarios.
Aphasia, a language disorder resulting from brain damage, requires accurate identification of specific aphasia types, such as Broca's and Wernicke's aphasia, for effective treatment. However, little attention has been paid to developing methods to detect different types of aphasia. Recognizing the importance of analyzing co-speech gestures for distinguish aphasia types, we propose a multimodal graph neural network for aphasia type detection using speech and corresponding gesture patterns. By learning the correlation between the speech and gesture modalities for each aphasia type, our model can generate textual representations sensitive to gesture information, leading to accurate aphasia type detection. Extensive experiments demonstrate the superiority of our approach over existing methods, achieving state-of-the-art results (F1 84.2\%). We also show that gesture features outperform acoustic features, highlighting the significance of gesture expression in detecting aphasia types. We provide the codes for reproducibility purposes.
Recurrent Neural Networks (RNNs) are useful in temporal sequence tasks. However, training RNNs involves dense matrix multiplications which require hardware that can support a large number of arithmetic operations and memory accesses. Implementing online training of RNNs on the edge calls for optimized algorithms for an efficient deployment on hardware. Inspired by the spiking neuron model, the Delta RNN exploits temporal sparsity during inference by skipping over the update of hidden states from those inactivated neurons whose change of activation across two timesteps is below a defined threshold. This work describes a training algorithm for Delta RNNs that exploits temporal sparsity in the backward propagation phase to reduce computational requirements for training on the edge. Due to the symmetric computation graphs of forward and backward propagation during training, the gradient computation of inactivated neurons can be skipped. Results show a reduction of $\sim$80% in matrix operations for training a 56k parameter Delta LSTM on the Fluent Speech Commands dataset with negligible accuracy loss. Logic simulations of a hardware accelerator designed for the training algorithm show 2-10X speedup in matrix computations for an activation sparsity range of 50%-90%. Additionally, we show that the proposed Delta RNN training will be useful for online incremental learning on edge devices with limited computing resources.
Speech is considered as a multi-modal process where hearing and vision are two fundamentals pillars. In fact, several studies have demonstrated that the robustness of Automatic Speech Recognition systems can be improved when audio and visual cues are combined to represent the nature of speech. In addition, Visual Speech Recognition, an open research problem whose purpose is to interpret speech by reading the lips of the speaker, has been a focus of interest in the last decades. Nevertheless, in order to estimate these systems in the currently Deep Learning era, large-scale databases are required. On the other hand, while most of these databases are dedicated to English, other languages lack sufficient resources. Thus, this paper presents a semi-automatically annotated audiovisual database to deal with unconstrained natural Spanish, providing 13 hours of data extracted from Spanish television. Furthermore, baseline results for both speaker-dependent and speaker-independent scenarios are reported using Hidden Markov Models, a traditional paradigm that has been widely used in the field of Speech Technologies.
We introduce a new zero resource code-switched speech benchmark designed to directly assess the code-switching capabilities of self-supervised speech encoders. We showcase a baseline system of language modeling on discrete units to demonstrate how the code-switching abilities of speech encoders can be assessed in a zero-resource manner. Our experiments encompass a variety of well-known speech encoders, including Wav2vec 2.0, HuBERT, XLSR, etc. We examine the impact of pre-training languages and model size on benchmark performance. Notably, though our results demonstrate that speech encoders with multilingual pre-training, exemplified by XLSR, outperform monolingual variants (Wav2vec 2.0, HuBERT) in code-switching scenarios, there is still substantial room for improvement in their code-switching linguistic abilities.
The prominence of a spoken word is the degree to which an average native listener perceives the word as salient or emphasized relative to its context. Speech prominence estimation is the process of assigning a numeric value to the prominence of each word in an utterance. These prominence labels are useful for linguistic analysis, as well as training automated systems to perform emphasis-controlled text-to-speech or emotion recognition. Manually annotating prominence is time-consuming and expensive, which motivates the development of automated methods for speech prominence estimation. However, developing such an automated system using machine-learning methods requires human-annotated training data. Using our system for acquiring such human annotations, we collect and open-source crowdsourced annotations of a portion of the LibriTTS dataset. We use these annotations as ground truth to train a neural speech prominence estimator that generalizes to unseen speakers, datasets, and speaking styles. We investigate design decisions for neural prominence estimation as well as how neural prominence estimation improves as a function of two key factors of annotation cost: dataset size and the number of annotations per utterance.