Automatic emotion recognition plays a key role in computer-human interaction as it has the potential to enrich the next-generation artificial intelligence with emotional intelligence. It finds applications in customer and/or representative behavior analysis in call centers, gaming, personal assistants, and social robots, to mention a few. Therefore, there has been an increasing demand to develop robust automatic methods to analyze and recognize the various emotions. In this paper, we propose a neural network-based emotion recognition framework that uses a late fusion of transfer-learned and fine-tuned models from speech and text modalities. More specifically, we i) adapt a residual network (ResNet) based model trained on a large-scale speaker recognition task using transfer learning along with a spectrogram augmentation approach to recognize emotions from speech, and ii) use a fine-tuned bidirectional encoder representations from transformers (BERT) based model to represent and recognize emotions from the text. The proposed system then combines the ResNet and BERT-based model scores using a late fusion strategy to further improve the emotion recognition performance. The proposed multimodal solution addresses the data scarcity limitation in emotion recognition using transfer learning, data augmentation, and fine-tuning, thereby improving the generalization performance of the emotion recognition models. We evaluate the effectiveness of our proposed multimodal approach on the interactive emotional dyadic motion capture (IEMOCAP) dataset. Experimental results indicate that both audio and text-based models improve the emotion recognition performance and that the proposed multimodal solution achieves state-of-the-art results on the IEMOCAP benchmark.
Online hate speech on social media has become a fast-growing problem in recent times. Nefarious groups have developed large content delivery networks across several main-stream (Twitter and Facebook) and fringe (Gab, 4chan, 8chan, etc.) outlets to deliver cascades of hate messages directed both at individuals and communities. Thus addressing these issues has become a top priority for large-scale social media outlets. Three key challenges in automated detection and classification of hateful content are the lack of clearly labeled data, evolving vocabulary and lexicon - hashtags, emojis, etc. - and the lack of baseline models for fringe outlets such as Gab. In this work, we propose a novel framework with three major contributions. (a) We engineer an ensemble of deep learning models that combines the strengths of state-of-the-art approaches, (b) we incorporate a tuning factor into this framework that leverages transfer learning to conduct automated hate speech classification on unlabeled datasets, like Gab, and (c) we develop a weak supervised learning methodology that allows our framework to train on unlabeled data. Our ensemble models achieve an 83% hate recall on the HON dataset, surpassing the performance of the state-of-the-art deep models. We demonstrate that weak supervised training in combination with classifier tuning significantly increases model performance on unlabeled data from Gab, achieving a hate recall of 67%.
The cloud-based speech recognition/API provides developers or enterprises an easy way to create speech-enabled features in their applications. However, sending audios about personal or company internal information to the cloud, raises concerns about the privacy and security issues. The recognition results generated in cloud may also reveal some sensitive information. This paper proposes a deep polynomial network (DPN) that can be applied to the encrypted speech as an acoustic model. It allows clients to send their data in an encrypted form to the cloud to ensure that their data remains confidential, at mean while the DPN can still make frame-level predictions over the encrypted speech and return them in encrypted form. One good property of the DPN is that it can be trained on unencrypted speech features in the traditional way. To keep the cloud away from the raw audio and recognition results, a cloud-local joint decoding framework is also proposed. We demonstrate the effectiveness of model and framework on the Switchboard and Cortana voice assistant tasks with small performance degradation and latency increased comparing with the traditional cloud-based DNNs.
End-to-end spoken language understanding (SLU) systems have many advantages over conventional pipeline systems, but collecting in-domain speech data to train an end-to-end system is costly and time consuming. One question arises from this: how to train an end-to-end SLU with limited amounts of data? Many researchers have explored approaches that make use of other related data resources, typically by pre-training parts of the model on high-resource speech recognition. In this paper, we suggest improving the generalization performance of SLU models with a non-standard learning algorithm, Reptile. Though Reptile was originally proposed for model-agnostic meta learning, we argue that it can also be used to directly learn a target task and result in better generalization than conventional gradient descent. In this work, we employ Reptile to the task of end-to-end spoken intent classification. Experiments on four datasets of different languages and domains show improvement of intent prediction accuracy, both when Reptile is used alone and used in addition to pre-training.
Recently neural architecture search(NAS) has been successfully used in image classification, natural language processing, and automatic speech recognition(ASR) tasks for finding the state-of-the-art(SOTA) architectures than those human-designed architectures. NAS can derive a SOTA and data-specific architecture over validation data from a pre-defined search space with a search algorithm. Inspired by the success of NAS in ASR tasks, we propose a NAS-based ASR framework containing one search space and one differentiable search algorithm called Differentiable Architecture Search(DARTS). Our search space follows the convolution-augmented transformer(Conformer) backbone, which is a more expressive ASR architecture than those used in existing NAS-based ASR frameworks. To improve the performance of our method, a regulation method called Dynamic Search Schedule(DSS) is employed. On a widely used Mandarin benchmark AISHELL-1, our best-searched architecture outperforms the baseline Conform model significantly with about 11% CER relative improvement, and our method is proved to be pretty efficient by the search cost comparisons.
The autoregressive (AR) models, such as attention-based encoder-decoder models and RNN-Transducer, have achieved great success in speech recognition. They predict the output sequence conditioned on the previous tokens and acoustic encoded states, which is inefficient on GPUs. The non-autoregressive (NAR) models can get rid of the temporal dependency between the output tokens and predict the entire output tokens in at least one step. However, the NAR model still faces two major problems. On the one hand, there is still a great gap in performance between the NAR models and the advanced AR models. On the other hand, it's difficult for most of the NAR models to train and converge. To address these two problems, we propose a new model named the two-step non-autoregressive transformer(TSNAT), which improves the performance and accelerating the convergence of the NAR model by learning prior knowledge from a parameters-sharing AR model. Furthermore, we introduce the two-stage method into the inference process, which improves the model performance greatly. All the experiments are conducted on a public Chinese mandarin dataset ASIEHLL-1. The results show that the TSNAT can achieve a competitive performance with the AR model and outperform many complicated NAR models.
Our paper focuses on the research of the possibility for speech recognition intelligent agents to assist the elderly and disabled people's lives, to improve their life quality by utilizing cutting-edge technologies. After researching the attitude of elderly and disabled people toward the household agent, we propose a design framework: ESNI(Exploration, Segmentation, Navigation, Instruction) that apply to mobile agent, achieve some functionalities such as processing human commands, picking up a specified object, and moving an object to another location. The agent starts the exploration in an unseen environment, stores each item's information in the grid cells to his memory and analyzes the corresponding features for each section. We divided our indoor environment into 6 sections: Kitchen, Living room, Bedroom, Studio, Bathroom, Balcony. The agent uses algorithms to assign sections for each grid cell then generates a navigation trajectory base on the section segmentation. When the user gives a command to the agent, feature words will be extracted and processed into a sequence of sub-tasks.
Automatic lyrics transcription (ALT), which can be regarded as automatic speech recognition (ASR) on singing voice, is an interesting and practical topic in academia and industry. ALT has not been well developed mainly due to the dearth of paired singing voice and lyrics datasets for model training. Considering that there is a large amount of ASR training data, a straightforward method is to leverage ASR data to enhance ALT training. However, the improvement is marginal when training the ALT system directly with ASR data, because of the gap between the singing voice and standard speech data which is rooted in music-specific acoustic characteristics in singing voice. In this paper, we propose PDAugment, a data augmentation method that adjusts pitch and duration of speech at syllable level under the guidance of music scores to help ALT training. Specifically, we adjust the pitch and duration of each syllable in natural speech to those of the corresponding note extracted from music scores, so as to narrow the gap between natural speech and singing voice. Experiments on DSing30 and Dali corpus show that the ALT system equipped with our PDAugment outperforms previous state-of-the-art systems by 5.9% and 18.1% WERs respectively, demonstrating the effectiveness of PDAugment for ALT.
The development of neural vocoders (NVs) has resulted in the high-quality and fast generation of waveforms. However, conventional NVs target a single sampling rate and require re-training when applied to different sampling rates. A suitable sampling rate varies from application to application due to the trade-off between speech quality and generation speed. In this study, we propose a method to handle multiple sampling rates in a single NV, called the MSR-NV. By generating waveforms step-by-step starting from a low sampling rate, MSR-NV can efficiently learn the characteristics of each frequency band and synthesize high-quality speech at multiple sampling rates. It can be regarded as an extension of the previously proposed NVs, and in this study, we extend the structure of Parallel WaveGAN (PWG). Experimental evaluation results demonstrate that the proposed method achieves remarkably higher subjective quality than the original PWG trained separately at 16, 24, and 48 kHz, without increasing the inference time. We also show that MSR-NV can leverage speech with lower sampling rates to further improve the quality of the synthetic speech.