Representations derived from models such as BERT (Bidirectional Encoder Representations from Transformers) and HuBERT (Hidden units BERT), have helped to achieve state-of-the-art performance in dimensional speech emotion recognition. Despite their large dimensionality, and even though these representations are not tailored for emotion recognition tasks, they are frequently used to train large speech emotion models with high memory and computational costs. In this work, we show that there exist lower-dimensional subspaces within the these pre-trained representational spaces that offer a reduction in downstream model complexity without sacrificing performance on emotion estimation. In addition, we model label uncertainty in the form of grader opinion variance, and demonstrate that such information can improve the models generalization capacity and robustness. Finally, we compare the robustness of the emotion models against acoustic degradations and observed that the reduced dimensional representations were able to retain the performance similar to the full-dimensional representations without significant regression in dimensional emotion performance.
Pre-trained model representations have demonstrated state-of-the-art performance in speech recognition, natural language processing, and other applications. Speech models, such as Bidirectional Encoder Representations from Transformers (BERT) and Hidden units BERT (HuBERT), have enabled generating lexical and acoustic representations to benefit speech recognition applications. We investigated the use of pre-trained model representations for estimating dimensional emotions, such as activation, valence, and dominance, from speech. We observed that while valence may rely heavily on lexical representations, activation and dominance rely mostly on acoustic information. In this work, we used multi-modal fusion representations from pre-trained models to generate state-of-the-art speech emotion estimation, and we showed a 100% and 30% relative improvement in concordance correlation coefficient (CCC) on valence estimation compared to standard acoustic and lexical baselines. Finally, we investigated the robustness of pre-trained model representations against noise and reverberation degradation and noticed that lexical and acoustic representations are impacted differently. We discovered that lexical representations are more robust to distortions compared to acoustic representations, and demonstrated that knowledge distillation from a multi-modal model helps to improve the noise-robustness of acoustic-based models.
Estimating dimensional emotions, such as activation, valence and dominance, from acoustic speech signals has been widely explored over the past few years. While accurate estimation of activation and dominance from speech seem to be possible, the same for valence remains challenging. Previous research has shown that the use of lexical information can improve valence estimation performance. Lexical information can be obtained from pre-trained acoustic models, where the learned representations can improve valence estimation from speech. We investigate the use of pre-trained model representations to improve valence estimation from acoustic speech signal. We also explore fusion of representations to improve emotion estimation across all three emotion dimensions: activation, valence and dominance. Additionally, we investigate if representations from pre-trained models can be distilled into models trained with low-level features, resulting in models with a less number of parameters. We show that fusion of pre-trained model embeddings result in a 79% relative improvement in concordance correlation coefficient CCC on valence estimation compared to standard acoustic feature baseline (mel-filterbank energies), while distillation from pre-trained model embeddings to lower-dimensional representations yielded a relative 12% improvement. Such performance gains were observed over two evaluation sets, indicating that our proposed architecture generalizes across those evaluation sets. We report new state-of-the-art "text-free" acoustic-only dimensional emotion estimation $CCC$ values on two MSP-Podcast evaluation sets.
Respiratory rate (RR) is a clinical metric used to assess overall health and physical fitness. An individual's RR can change from their baseline due to chronic illness symptoms (e.g., asthma, congestive heart failure), acute illness (e.g., breathlessness due to infection), and over the course of the day due to physical exhaustion during heightened exertion. Remote estimation of RR can offer a cost-effective method to track disease progression and cardio-respiratory fitness over time. This work investigates a model-driven approach to estimate RR from short audio segments obtained after physical exertion in healthy adults. Data was collected from 21 individuals using microphone-enabled, near-field headphones before, during, and after strenuous exercise. RR was manually annotated by counting perceived inhalations and exhalations. A multi-task Long-Short Term Memory (LSTM) network with convolutional layers was implemented to process mel-filterbank energies, estimate RR in varying background noise conditions, and predict heavy breathing, indicated by an RR of more than 25 breaths per minute. The multi-task model performs both classification and regression tasks and leverages a mixture of loss functions. It was observed that RR can be estimated with a concordance correlation coefficient (CCC) of 0.76 and a mean squared error (MSE) of 0.2, demonstrating that audio can be a viable signal for approximating RR.
Dysfluencies and variations in speech pronunciation can severely degrade speech recognition performance, and for many individuals with moderate-to-severe speech disorders, voice operated systems do not work. Current speech recognition systems are trained primarily with data from fluent speakers and as a consequence do not generalize well to speech with dysfluencies such as sound or word repetitions, sound prolongations, or audible blocks. The focus of this work is on quantitative analysis of a consumer speech recognition system on individuals who stutter and production-oriented approaches for improving performance for common voice assistant tasks (i.e., "what is the weather?"). At baseline, this system introduces a significant number of insertion and substitution errors resulting in intended speech Word Error Rates (isWER) that are 13.64\% worse (absolute) for individuals with fluency disorders. We show that by simply tuning the decoding parameters in an existing hybrid speech recognition system one can improve isWER by 24\% (relative) for individuals with fluency disorders. Tuning these parameters translates to 3.6\% better domain recognition and 1.7\% better intent recognition relative to the default setup for the 18 study participants across all stuttering severities.
The ability to automatically detect stuttering events in speech could help speech pathologists track an individual's fluency over time or help improve speech recognition systems for people with atypical speech patterns. Despite increasing interest in this area, existing public datasets are too small to build generalizable dysfluency detection systems and lack sufficient annotations. In this work, we introduce Stuttering Events in Podcasts (SEP-28k), a dataset containing over 28k clips labeled with five event types including blocks, prolongations, sound repetitions, word repetitions, and interjections. Audio comes from public podcasts largely consisting of people who stutter interviewing other people who stutter. We benchmark a set of acoustic models on SEP-28k and the public FluencyBank dataset and highlight how simply increasing the amount of training data improves relative detection performance by 28\% and 24\% F1 on each. Annotations from over 32k clips across both datasets will be publicly released.
Emotion plays an essential role in human-to-human communication, enabling us to convey feelings such as happiness, frustration, and sincerity. While modern speech technologies rely heavily on speech recognition and natural language understanding for speech content understanding, the investigation of vocal expression is increasingly gaining attention. Key considerations for building robust emotion models include characterizing and improving the extent to which a model, given its training data distribution, is able to generalize to unseen data conditions. This work investigated a long-shot-term memory (LSTM) network and a time convolution - LSTM (TC-LSTM) to detect primitive emotion attributes such as valence, arousal, and dominance, from speech. It was observed that training with multiple datasets and using robust features improved the concordance correlation coefficient (CCC) for valence, by 30\% with respect to the baseline system. Additionally, this work investigated how emotion primitives can be used to detect categorical emotions such as happiness, disgust, contempt, anger, and surprise from neutral speech, and results indicated that arousal, followed by dominance was a better detector of such emotions.
Unseen or out-of-domain data can seriously degrade the performance of a neural network model, indicating the model's failure to generalize to unseen data. Neural net pruning can not only help to reduce a model's size but can improve the model's generalization capacity as well. Pruning approaches look for low-salient neurons that are less contributive to a model's decision and hence can be removed from the model. This work investigates if pruning approaches are successful in detecting neurons that are either high-salient (mostly active or hyper) or low-salient (barely active or hypo), and whether removal of such neurons can help to improve the model's generalization capacity. Traditional blind adaptation techniques update either the whole or a subset of layers, but have never explored selectively updating individual neurons across one or more layers. Focusing on the fully connected layers of a convolutional neural network (CNN), this work shows that it may be possible to selectively adapt certain neurons (consisting of the hyper and the hypo neurons) first, followed by a full-network fine tuning. Using the task of automatic speech recognition, this work demonstrates how the removal of hyper and hypo neurons from a model can improve the model's performance on out-of-domain speech data and how selective neuron adaptation can ensure improved performance when compared to traditional blind model adaptation.
Millions of people reach out to digital assistants such as Siri every day, asking for information, making phone calls, seeking assistance, and much more. The expectation is that such assistants should understand the intent of the users query. Detecting the intent of a query from a short, isolated utterance is a difficult task. Intent cannot always be obtained from speech-recognized transcriptions. A transcription driven approach can interpret what has been said but fails to acknowledge how it has been said, and as a consequence, may ignore the expression present in the voice. Our work investigates whether a system can reliably detect vocal expression in queries using acoustic and paralinguistic embedding. Results show that the proposed method offers a relative equal error rate (EER) decrease of 60% compared to a bag-of-word based system, corroborating that expression is significantly represented by vocal attributes, rather than being purely lexical. Addition of emotion embedding helped to reduce the EER by 30% relative to the acoustic embedding, demonstrating the relevance of emotion in expressive voice.
The rapid population aging has stimulated the development of assistive devices that provide personalized medical support to the needies suffering from various etiologies. One prominent clinical application is a computer-assisted speech training system which enables personalized speech therapy to patients impaired by communicative disorders in the patient's home environment. Such a system relies on the robust automatic speech recognition (ASR) technology to be able to provide accurate articulation feedback. With the long-term aim of developing off-the-shelf ASR systems that can be incorporated in clinical context without prior speaker information, we compare the ASR performance of speaker-independent bottleneck and articulatory features on dysarthric speech used in conjunction with dedicated neural network-based acoustic models that have been shown to be robust against spectrotemporal deviations. We report ASR performance of these systems on two dysarthric speech datasets of different characteristics to quantify the achieved performance gains. Despite the remaining performance gap between the dysarthric and normal speech, significant improvements have been reported on both datasets using speaker-independent ASR architectures.