Electrical Engineering And Systems Science

Multimodal emotion recognition from speech is an important area in affective computing. Fusing multiple data modalities and learning representations with limited amounts of labeled data is a challenging task. In this paper, we explore the use of modality-specific "BERT-like" pretrained Self Supervised Learning (SSL) architectures to represent both speech and text modalities for the task of multimodal speech emotion recognition. By conducting experiments on three publicly available datasets (IEMOCAP, CMU-MOSEI, and CMU-MOSI), we show that jointly fine-tuning "BERT-like" SSL architectures achieve state-of-the-art (SOTA) results. We also evaluate two methods of fusing speech and text modalities and show that a simple fusion mechanism can outperform more complex ones when using SSL models that have similar architectural properties to BERT.

Deep learning techniques have considerably improved speech processing in recent years. Speech representations extracted by deep learning models are being used in a wide range of tasks such as speech recognition, speaker recognition, and speech emotion recognition. Attention models play an important role in improving deep learning models. However current attention mechanisms are unable to attend to fine-grained information items. In this paper we propose the novel Fine-grained Early Frequency Attention (FEFA) for speech signals. This model is capable of focusing on information items as small as frequency bins. We evaluate the proposed model on two popular tasks of speaker recognition and speech emotion recognition. Two widely used public datasets, VoxCeleb and IEMOCAP, are used for our experiments. The model is implemented on top of several prominent deep models as backbone networks to evaluate its impact on performance compared to the original networks and other related work. Our experiments show that by adding FEFA to different CNN architectures, performance is consistently improved by substantial margins, even setting a new state-of-the-art for the speaker recognition task. We also tested our model against different levels of added noise showing improvements in robustness and less sensitivity compared to the backbone networks.

We present KoSpeech, an open-source software, which is modular and extensible end-to-end Korean automatic speech recognition (ASR) toolkit based on the deep learning library PyTorch. Several automatic speech recognition open-source toolkits have been released, but all of them deal with non-Korean languages, such as English (e.g. ESPnet, Espresso). Although AI Hub opened 1,000 hours of Korean speech corpus known as KsponSpeech, there is no established preprocessing method and baseline model to compare model performances. Therefore, we propose preprocessing methods for KsponSpeech corpus and a baseline model for benchmarks. Our baseline model is based on Listen, Attend and Spell (LAS) architecture and ables to customize various training hyperparameters conveniently. By KoSpeech, we hope this could be a guideline for those who research Korean speech recognition. Our baseline model achieved 10.31% character error rate (CER) at KsponSpeech corpus only with the acoustic model. Our source code is available here.

Speech synthesis (text to speech, TTS) and recognition (automatic speech recognition, ASR) are important speech tasks, and require a large amount of text and speech pairs for model training. However, there are more than 6,000 languages in the world and most languages are lack of speech training data, which poses significant challenges when building TTS and ASR systems for extremely low-resource languages. In this paper, we develop LRSpeech, a TTS and ASR system under the extremely low-resource setting, which can support rare languages with low data cost. LRSpeech consists of three key techniques: 1) pre-training on rich-resource languages and fine-tuning on low-resource languages; 2) dual transformation between TTS and ASR to iteratively boost the accuracy of each other; 3) knowledge distillation to customize the TTS model on a high-quality target-speaker voice and improve the ASR model on multiple voices. We conduct experiments on an experimental language (English) and a truly low-resource language (Lithuanian) to verify the effectiveness of LRSpeech. Experimental results show that LRSpeech 1) achieves high quality for TTS in terms of both intelligibility (more than 98% intelligibility rate) and naturalness (above 3.5 mean opinion score (MOS)) of the synthesized speech, which satisfy the requirements for industrial deployment, 2) achieves promising recognition accuracy for ASR, and 3) last but not least, uses extremely low-resource training data. We also conduct comprehensive analyses on LRSpeech with different amounts of data resources, and provide valuable insights and guidances for industrial deployment. We are currently deploying LRSpeech into a commercialized cloud speech service to support TTS on more rare languages.

Automated speech recognition coverage of the world's languages continues to expand. However, standard phoneme based systems require handcrafted lexicons that are difficult and expensive to obtain. To address this problem, we propose a training methodology for a grapheme-based speech recognizer that can be trained in a purely data-driven fashion. Built with LSTM networks and trained with the cross-entropy loss, the grapheme-output acoustic models we study are also extremely practical for real-world applications as they can be decoded with conventional ASR stack components such as language models and FST decoders, and produce good quality audio-to-grapheme alignments that are useful in many speech applications. We show that the grapheme models are competitive in WER with their phoneme-output counterparts when trained on large datasets, with the advantage that grapheme models do not require explicit linguistic knowledge as an input. We further compare the alignments generated by the phoneme and grapheme models to demonstrate the quality of the pronunciations learnt by them using four Indian languages that vary linguistically in spoken and written forms.

In speech recognition problems, data scarcity often poses an issue due to the willingness of humans to provide large amounts of data for learning and classification. In this work, we take a set of 5 spoken Harvard sentences from 7 subjects and consider their MFCC attributes. Using character level LSTMs (supervised learning) and OpenAI's attention-based GPT-2 models, synthetic MFCCs are generated by learning from the data provided on a per-subject basis. A neural network is trained to classify the data against a large dataset of Flickr8k speakers and is then compared to a transfer learning network performing the same task but with an initial weight distribution dictated by learning from the synthetic data generated by the two models. The best result for all of the 7 subjects were networks that had been exposed to synthetic data, the model pre-trained with LSTM-produced data achieved the best result 3 times and the GPT-2 equivalent 5 times (since one subject had their best result from both models at a draw). Through these results, we argue that speaker classification can be improved by utilising a small amount of user data but with exposure to synthetically-generated MFCCs which then allow the networks to achieve near maximum classification scores.

In this paper, we propose a novel conditional convolution network, named location-variable convolution, to model the dependencies of the waveform sequence. Different from the use of unified convolution kernels in WaveNet to capture the dependencies of arbitrary waveform, the location-variable convolution uses convolution kernels with different coefficients to perform convolution operations on different waveform intervals, where the coefficients of kernels is predicted according to conditioning acoustic features, such as Mel-spectrograms. Based on location-variable convolutions, we design LVCNet for waveform generation, and apply it in Parallel WaveGAN to design more efficient vocoder. Experiments on the LJSpeech dataset show that our proposed model achieves a four-fold increase in synthesis speed compared to the original Parallel WaveGAN without any degradation in sound quality, which verifies the effectiveness of location-variable convolutions.

End-to-end Spoken Language Understanding (SLU) models are made increasingly large and complex to achieve the state-ofthe-art accuracy. However, the increased complexity of a model can also introduce high risk of over-fitting, which is a major challenge in SLU tasks due to the limitation of available data. In this paper, we propose an attention-based SLU model together with three encoder enhancement strategies to overcome data sparsity challenge. The first strategy focuses on the transferlearning approach to improve feature extraction capability of the encoder. It is implemented by pre-training the encoder component with a quantity of Automatic Speech Recognition annotated data relying on the standard Transformer architecture and then fine-tuning the SLU model with a small amount of target labelled data. The second strategy adopts multitask learning strategy, the SLU model integrates the speech recognition model by sharing the same underlying encoder, such that improving robustness and generalization ability. The third strategy, learning from Component Fusion (CF) idea, involves a Bidirectional Encoder Representation from Transformer (BERT) model and aims to boost the capability of the decoder with an auxiliary network. It hence reduces the risk of over-fitting and augments the ability of the underlying encoder, indirectly. Experiments on the FluentAI dataset show that cross-language transfer learning and multi-task strategies have been improved by up to 4:52% and 3:89% respectively, compared to the baseline.

Despite the growing interest for expressive speech synthesis, synthesis of nonverbal expressions is an under-explored area. In this paper we propose an audio laughter synthesis system based on a sequence-to-sequence TTS synthesis system. We leverage transfer learning by training a deep learning model to learn to generate both speech and laughs from annotations. We evaluate our model with a listening test, comparing its performance to an HMM-based laughter synthesis one and assess that it reaches higher perceived naturalness. Our solution is a first step towards a TTS system that would be able to synthesize speech with a control on amusement level with laughter integration.

In this paper, we explore the neural architecture search (NAS) for automatic speech recognition (ASR) systems. With reference to the previous works in the computer vision field, the transferability of the searched architecture is the main focus of our work. The architecture search is conducted on the small proxy dataset, and then the evaluation network, constructed with the searched architecture, is evaluated on the large dataset. Especially, we propose a revised search space for speech recognition tasks which theoretically facilitates the search algorithm to explore the architectures with low complexity. Extensive experiments show that: (i) the architecture searched on the small proxy dataset can be transferred to the large dataset for the speech recognition tasks. (ii) the architecture learned in the revised search space can greatly reduce the computational overhead and GPU memory usage with mild performance degradation. (iii) the searched architecture can achieve more than 20% and 15% (average on the four test sets) relative improvements respectively on the AISHELL-2 dataset and the large (10k hours) dataset, compared with our best hand-designed DFSMN-SAN architecture. To the best of our knowledge, this is the first report of NAS results with large scale dataset (up to 10K hours), indicating the promising application of NAS to industrial ASR systems.