Lirong Dai

Ziqiang Zhang, Sanyuan Chen, Long Zhou et al. · microsoft-research

How to boost speech pre-training with textual data is an unsolved problem due to the fact that speech and text are very different modalities with distinct characteristics. In this paper, we propose a cross-modal Speech and Language Model (SpeechLM) to explicitly align speech and text pre-training with a pre-defined unified discrete representation. Specifically, we introduce two alternative discrete tokenizers to bridge the speech and text modalities, including phoneme-unit and hidden-unit tokenizers, which can be trained using a small amount of paired speech-text data. Based on the trained tokenizers, we convert the unlabeled speech and text data into tokens of phoneme units or hidden units. The pre-training objective is designed to unify the speech and the text into the same discrete semantic space with a unified Transformer network. We evaluate SpeechLM on various spoken language processing tasks including speech recognition, speech translation, and universal representation evaluation framework SUPERB, demonstrating significant improvements on content-related tasks. Code and models are available at https://aka.ms/SpeechLM.

Ziqiang Zhang, Long Zhou, Junyi Ao et al. · microsoft-research

The rapid development of single-modal pre-training has prompted researchers to pay more attention to cross-modal pre-training methods. In this paper, we propose a unified-modal speech-unit-text pre-training model, SpeechUT, to connect the representations of a speech encoder and a text decoder with a shared unit encoder. Leveraging hidden-unit as an interface to align speech and text, we can decompose the speech-to-text model into a speech-to-unit model and a unit-to-text model, which can be jointly pre-trained with unpaired speech and text data respectively. Our proposed SpeechUT is fine-tuned and evaluated on automatic speech recognition (ASR) and speech translation (ST) tasks. Experimental results show that SpeechUT gets substantial improvements over strong baselines, and achieves state-of-the-art performance on both the LibriSpeech ASR and MuST-C ST tasks. To better understand the proposed SpeechUT, detailed analyses are conducted. The code and pre-trained models are available at https://aka.ms/SpeechUT.

Qiushi Zhu, Long Zhou, Ziqiang Zhang et al. · microsoft-research

Although speech is a simple and effective way for humans to communicate with the outside world, a more realistic speech interaction contains multimodal information, e.g., vision, text. How to design a unified framework to integrate different modal information and leverage different resources (e.g., visual-audio pairs, audio-text pairs, unlabeled speech, and unlabeled text) to facilitate speech representation learning was not well explored. In this paper, we propose a unified cross-modal representation learning framework VATLM (Visual-Audio-Text Language Model). The proposed VATLM employs a unified backbone network to model the modality-independent information and utilizes three simple modality-dependent modules to preprocess visual, speech, and text inputs. In order to integrate these three modalities into one shared semantic space, VATLM is optimized with a masked prediction task of unified tokens, given by our proposed unified tokenizer. We evaluate the pre-trained VATLM on audio-visual related downstream tasks, including audio-visual speech recognition (AVSR), visual speech recognition (VSR) tasks. Results show that the proposed VATLM outperforms previous the state-of-the-art models, such as audio-visual pre-trained AV-HuBERT model, and analysis also demonstrates that VATLM is capable of aligning different modalities into the same space. To facilitate future research, we release the code and pre-trained models at https://aka.ms/vatlm.

Dianhai Yu, Liang Shen, Hongxiang Hao et al.

While modern internet services, such as chatbots, search engines, and online advertising, demand the use of large-scale deep neural networks (DNNs), distributed training and inference over heterogeneous computing systems are desired to facilitate these DNN models. Mixture-of-Experts (MoE) is one the most common strategies to lower the cost of training subject to the overall size of models/data through gating and parallelism in a divide-and-conquer fashion. While DeepSpeed has made efforts in carrying out large-scale MoE training over heterogeneous infrastructures, the efficiency of training and inference could be further improved from several system aspects, including load balancing, communication/computation efficiency, and memory footprint limits. In this work, we present a novel MoESys that boosts efficiency in both large-scale training and inference. Specifically, in the training procedure, the proposed MoESys adopts an Elastic MoE training strategy with 2D prefetch and Fusion communication over Hierarchical storage, so as to enjoy efficient parallelisms. For scalable inference in a single node, especially when the model size is larger than GPU memory, MoESys builds the CPU-GPU memory jointly into a ring of sections to load the model, and executes the computation tasks across the memory sections in a round-robin manner for efficient inference. We carried out extensive experiments to evaluate MoESys, where MoESys successfully trains a Unified Feature Optimization (UFO) model with a Sparsely-Gated Mixture-of-Experts model of 12B parameters in 8 days on 48 A100 GPU cards. The comparison against the state-of-the-art shows that MoESys outperformed DeepSpeed with 33% higher throughput (tokens per second) in training and 13% higher throughput in inference in general. Particularly, under unbalanced MoE Tasks, e.g., UFO, MoESys achieved 64% higher throughput with 18% lower memory footprints.

Jinshui Hu, Chenyu Liu, Qiandong Yan et al.

Recent works have shown huge success of deep learning models for common in vocabulary (IV) scene text recognition. However, in real-world scenarios, out-of-vocabulary (OOV) words are of great importance and SOTA recognition models usually perform poorly on OOV settings. Inspired by the intuition that the learned language prior have limited OOV preformence, we design a framework named Vision Language Adaptive Mutual Decoder (VLAMD) to tackle OOV problems partly. VLAMD consists of three main conponents. Firstly, we build an attention based LSTM decoder with two adaptively merged visual-only modules, yields a vision-language balanced main branch. Secondly, we add an auxiliary query based autoregressive transformer decoding head for common visual and language prior representation learning. Finally, we couple these two designs with bidirectional training for more diverse language modeling, and do mutual sequential decoding to get robuster results. Our approach achieved 70.31\% and 59.61\% word accuracy on IV+OOV and OOV settings respectively on Cropped Word Recognition Task of OOV-ST Challenge at ECCV 2022 TiE Workshop, where we got 1st place on both settings.

Jian Zhu, Xin Zou, Jun Sun et al.

In recent years, the advancement of Graph Neural Networks (GNNs) has significantly propelled progress in Multi-View Clustering (MVC). However, existing methods face the problem of coarse-grained graph fusion. Specifically, current approaches typically generate a separate graph structure for each view and then perform weighted fusion of graph structures at the view level, which is a relatively rough strategy. To address this limitation, we present a novel Mixture of Ego-Graphs Contrastive Representation Learning (MoEGCL). It mainly consists of two modules. In particular, we propose an innovative Mixture of Ego-Graphs Fusion (MoEGF), which constructs ego graphs and utilizes a Mixture-of-Experts network to implement fine-grained fusion of ego graphs at the sample level, rather than the conventional view-level fusion. Additionally, we present the Ego Graph Contrastive Learning (EGCL) module to align the fused representation with the view-specific representation. The EGCL module enhances the representation similarity of samples from the same cluster, not merely from the same sample, further boosting fine-grained graph representation. Extensive experiments demonstrate that MoEGCL achieves state-of-the-art results in deep multi-view clustering tasks. The source code is publicly available at https://github.com/HackerHyper/MoEGCL.

Yanfeng Shi, Pengfei Cai, Jun Liu et al.

Large Audio-Language Models (LALMs) enable general audio understanding and demonstrate remarkable performance across various audio tasks. However, these models still face challenges in temporal perception (e.g., inferring event onset and offset), leading to limited utility in fine-grained scenarios. To address this issue, we propose Audio-Side Time Prompt and leverage Reinforcement Learning (RL) to develop the TimePro-RL framework for fine-grained temporal perception. Specifically, we encode timestamps as embeddings and interleave them within the audio feature sequence as temporal coordinates to prompt the model. Furthermore, we introduce RL following Supervised Fine-Tuning (SFT) to directly optimize temporal alignment performance. Experiments demonstrate that TimePro-RL achieves significant performance gains across a range of audio temporal tasks, such as audio grounding, sound event detection, and dense audio captioning, validating its robust effectiveness.

Mengzhi Wang, Shifu Xiong, Genshun Wan et al.

Contextual-LAS (CLAS) has been shown effective in improving Automatic Speech Recognition (ASR) of rare words. It relies on phrase-level contextual modeling and attention-based relevance scoring without explicit contextual constraint which lead to insufficient use of contextual information. In this work, we propose deep CLAS to use contextual information better. We introduce bias loss forcing model to focus on contextual information. The query of bias attention is also enriched to improve the accuracy of the bias attention score. To get fine-grained contextual information, we replace phrase-level encoding with character-level encoding and encode contextual information with conformer rather than LSTM. Moreover, we directly use the bias attention score to correct the output probability distribution of the model. Experiments using the public AISHELL-1 and AISHELL-NER. On AISHELL-1, compared to CLAS baselines, deep CLAS obtains a 65.78% relative recall and a 53.49% relative F1-score increase in the named entity recognition scene.

Junyi Ao, Ziqiang Zhang, Long Zhou et al.

This paper studies a novel pre-training technique with unpaired speech data, Speech2C, for encoder-decoder based automatic speech recognition (ASR). Within a multi-task learning framework, we introduce two pre-training tasks for the encoder-decoder network using acoustic units, i.e., pseudo codes, derived from an offline clustering model. One is to predict the pseudo codes via masked language modeling in encoder output, like HuBERT model, while the other lets the decoder learn to reconstruct pseudo codes autoregressively instead of generating textual scripts. In this way, the decoder learns to reconstruct original speech information with codes before learning to generate correct text. Comprehensive experiments on the LibriSpeech corpus show that the proposed Speech2C can relatively reduce the word error rate (WER) by 19.2% over the method without decoder pre-training, and also outperforms significantly the state-of-the-art wav2vec 2.0 and HuBERT on fine-tuning subsets of 10h and 100h. We release our code and model at https://github.com/microsoft/SpeechT5/tree/main/Speech2C.

Jianwei Cui, Yu Gu, Chao Weng et al.

This paper presents an advanced end-to-end singing voice synthesis (SVS) system based on the source-filter mechanism that directly translates lyrical and melodic cues into expressive and high-fidelity human-like singing. Similarly to VISinger 2, the proposed system also utilizes training paradigms evolved from VITS and incorporates elements like the fundamental pitch (F0) predictor and waveform generation decoder. To address the issue that the coupling of mel-spectrogram features with F0 information may introduce errors during F0 prediction, we consider two strategies. Firstly, we leverage mel-cepstrum (mcep) features to decouple the intertwined mel-spectrogram and F0 characteristics. Secondly, inspired by the neural source-filter models, we introduce source excitation signals as the representation of F0 in the SVS system, aiming to capture pitch nuances more accurately. Meanwhile, differentiable mcep and F0 losses are employed as the waveform decoder supervision to fortify the prediction accuracy of speech envelope and pitch in the generated speech. Experiments on the Opencpop dataset demonstrate efficacy of the proposed model in synthesis quality and intonation accuracy.

Dan Liu, Mengge Du, Xiaoxi Li et al.

This paper describes USTC-NELSLIP's submissions to the IWSLT2021 Simultaneous Speech Translation task. We proposed a novel simultaneous translation model, Cross Attention Augmented Transducer (CAAT), which extends conventional RNN-T to sequence-to-sequence tasks without monotonic constraints, e.g., simultaneous translation. Experiments on speech-to-text (S2T) and text-to-text (T2T) simultaneous translation tasks shows CAAT achieves better quality-latency trade-offs compared to \textit{wait-k}, one of the previous state-of-the-art approaches. Based on CAAT architecture and data augmentation, we build S2T and T2T simultaneous translation systems in this evaluation campaign. Compared to last year's optimal systems, our S2T simultaneous translation system improves by an average of 11.3 BLEU for all latency regimes, and our T2T simultaneous translation system improves by an average of 4.6 BLEU.

Liangfa Wei, Jie Zhang, Junfeng Hou et al.

Audio-visual information fusion enables a performance improvement in speech recognition performed in complex acoustic scenarios, e.g., noisy environments. It is required to explore an effective audio-visual fusion strategy for audiovisual alignment and modality reliability. Different from the previous end-to-end approaches where the audio-visual fusion is performed after encoding each modality, in this paper we propose to integrate an attentive fusion block into the encoding process. It is shown that the proposed audio-visual fusion method in the encoder module can enrich audio-visual representations, as the relevance between the two modalities is leveraged. In line with the transformer-based architecture, we implement the embedded fusion block using a multi-head attention based audiovisual fusion with one-way or two-way interactions. The proposed method can sufficiently combine the two streams and weaken the over-reliance on the audio modality. Experiments on the LRS3-TED dataset demonstrate that the proposed method can increase the recognition rate by 0.55%, 4.51% and 4.61% on average under the clean, seen and unseen noise conditions, respectively, compared to the state-of-the-art approach.

Fenglin Ding, Wu Guo, Lirong Dai et al.

Batch normalization (BN) is an effective method to accelerate model training and improve the generalization performance of neural networks. In this paper, we propose an improved batch normalization technique called attentive batch normalization (ABN) in Long Short Term Memory (LSTM) based acoustic modeling for automatic speech recognition (ASR). In the proposed method, an auxiliary network is used to dynamically generate the scaling and shifting parameters in batch normalization, and attention mechanisms are introduced to improve their regularized performance. Furthermore, two schemes, frame-level and utterance-level ABN, are investigated. We evaluate our proposed methods on Mandarin and Uyghur ASR tasks, respectively. The experimental results show that the proposed ABN greatly improves the performance of batch normalization in terms of transcription accuracy for both languages.

Lanhua You, Wu Guo, Lirong Dai et al.

In this paper, gating mechanisms are applied in deep neural network (DNN) training for x-vector-based text-independent speaker verification. First, a gated convolution neural network (GCNN) is employed for modeling the frame-level embedding layers. Compared with the time-delay DNN (TDNN), the GCNN can obtain more expressive frame-level representations through carefully designed memory cell and gating mechanisms. Moreover, we propose a novel gated-attention statistics pooling strategy in which the attention scores are shared with the output gate. The gated-attention statistics pooling combines both gating and attention mechanisms into one framework; therefore, we can capture more useful information in the temporal pooling layer. Experiments are carried out using the NIST SRE16 and SRE18 evaluation datasets. The results demonstrate the effectiveness of the GCNN and show that the proposed gated-attention statistics pooling can further improve the performance.

Lanhua You, Wu Guo, Lirong Dai et al.

The x-vector based deep neural network (DNN) embedding systems have demonstrated effectiveness for text-independent speaker verification. This paper presents a multi-task learning architecture for training the speaker embedding DNN with the primary task of classifying the target speakers, and the auxiliary task of reconstructing the first- and higher-order statistics of the original input utterance. The proposed training strategy aggregates both the supervised and unsupervised learning into one framework to make the speaker embeddings more discriminative and robust. Experiments are carried out using the NIST SRE16 evaluation dataset and the VOiCES dataset. The results demonstrate that our proposed method outperforms the original x-vector approach with very low additional complexity added.

Yaming Liu, Jian Tang, Yan Song et al.

Polyphonic sound event detection (polyphonic SED) is an interesting but challenging task due to the concurrence of multiple sound events. Recently, SED methods based on convolutional neural networks (CNN) and recurrent neural networks (RNN) have shown promising performance. Generally, CNN are designed for local feature extraction while RNN are used to model the temporal dependency among these local features. Despite their success, it is still insufficient for existing deep learning techniques to separate individual sound event from their mixture, largely due to the overlapping characteristic of features. Motivated by the success of Capsule Networks (CapsNet), we propose a more suitable capsule based approach for polyphonic SED. Specifically, several capsule layers are designed to effectively select representative frequency bands for each individual sound event. The temporal dependency of capsule's outputs is then modeled by a RNN. And a dynamic threshold method is proposed for making the final decision based on RNN outputs. Experiments on the TUT-SED Synthetic 2016 dataset show that the proposed approach obtains an F1-score of 68.8% and an error rate of 0.45, outperforming the previous state-of-the-art method of 66.4% and 0.48, respectively.

Shiliang Zhang, Ming Lei, Zhijie Yan et al.

In this paper, we present an improved feedforward sequential memory networks (FSMN) architecture, namely Deep-FSMN (DFSMN), by introducing skip connections between memory blocks in adjacent layers. These skip connections enable the information flow across different layers and thus alleviate the gradient vanishing problem when building very deep structure. As a result, DFSMN significantly benefits from these skip connections and deep structure. We have compared the performance of DFSMN to BLSTM both with and without lower frame rate (LFR) on several large speech recognition tasks, including English and Mandarin. Experimental results shown that DFSMN can consistently outperform BLSTM with dramatic gain, especially trained with LFR using CD-Phone as modeling units. In the 2000 hours Fisher (FSH) task, the proposed DFSMN can achieve a word error rate of 9.4% by purely using the cross-entropy criterion and decoding with a 3-gram language model, which achieves a 1.5% absolute improvement compared to the BLSTM. In a 20000 hours Mandarin recognition task, the LFR trained DFSMN can achieve more than 20% relative improvement compared to the LFR trained BLSTM. Moreover, we can easily design the lookahead filter order of the memory blocks in DFSMN to control the latency for real-time applications.

Jianshu Zhang, Yixing Zhu, Jun Du et al.

Recently, great progress has been made for online handwritten Chinese character recognition due to the emergence of deep learning techniques. However, previous research mostly treated each Chinese character as one class without explicitly considering its inherent structure, namely the radical components with complicated geometry. In this study, we propose a novel trajectory-based radical analysis network (TRAN) to firstly identify radicals and analyze two-dimensional structures among radicals simultaneously, then recognize Chinese characters by generating captions of them based on the analysis of their internal radicals. The proposed TRAN employs recurrent neural networks (RNNs) as both an encoder and a decoder. The RNN encoder makes full use of online information by directly transforming handwriting trajectory into high-level features. The RNN decoder aims at generating the caption by detecting radicals and spatial structures through an attention model. The manner of treating a Chinese character as a two-dimensional composition of radicals can reduce the size of vocabulary and enable TRAN to possess the capability of recognizing unseen Chinese character classes, only if the corresponding radicals have been seen. Evaluated on CASIA-OLHWDB database, the proposed approach significantly outperforms the state-of-the-art whole-character modeling approach with a relative character error rate (CER) reduction of 10%. Meanwhile, for the case of recognition of 500 unseen Chinese characters, TRAN can achieve a character accuracy of about 60% while the traditional whole-character method has no capability to handle them.

Jianshu Zhang, Jun Du, Lirong Dai

Handwritten mathematical expression recognition is a challenging problem due to the complicated two-dimensional structures, ambiguous handwriting input and variant scales of handwritten math symbols. To settle this problem, we utilize the attention based encoder-decoder model that recognizes mathematical expression images from two-dimensional layouts to one-dimensional LaTeX strings. We improve the encoder by employing densely connected convolutional networks as they can strengthen feature extraction and facilitate gradient propagation especially on a small training set. We also present a novel multi-scale attention model which is employed to deal with the recognition of math symbols in different scales and save the fine-grained details that will be dropped by pooling operations. Validated on the CROHME competition task, the proposed method significantly outperforms the state-of-the-art methods with an expression recognition accuracy of 52.8% on CROHME 2014 and 50.1% on CROHME 2016, by only using the official training dataset.

Jianshu Zhang, Jun Du, Lirong Dai

In this study, we present a novel end-to-end approach based on the encoder-decoder framework with the attention mechanism for online handwritten mathematical expression recognition (OHMER). First, the input two-dimensional ink trajectory information of handwritten expression is encoded via the gated recurrent unit based recurrent neural network (GRU-RNN). Then the decoder is also implemented by the GRU-RNN with a coverage-based attention model. The proposed approach can simultaneously accomplish the symbol recognition and structural analysis to output a character sequence in LaTeX format. Validated on the CROHME 2014 competition task, our approach significantly outperforms the state-of-the-art with an expression recognition accuracy of 52.43% by only using the official training dataset. Furthermore, the alignments between the input trajectories of handwritten expressions and the output LaTeX sequences are visualized by the attention mechanism to show the effectiveness of the proposed method.

Jianshu Zhang, Yixing Zhu, Jun Du et al.

Chinese characters have a huge set of character categories, more than 20,000 and the number is still increasing as more and more novel characters continue being created. However, the enormous characters can be decomposed into a compact set of about 500 fundamental and structural radicals. This paper introduces a novel radical analysis network (RAN) to recognize printed Chinese characters by identifying radicals and analyzing two-dimensional spatial structures among them. The proposed RAN first extracts visual features from input by employing convolutional neural networks as an encoder. Then a decoder based on recurrent neural networks is employed, aiming at generating captions of Chinese characters by detecting radicals and two-dimensional structures through a spatial attention mechanism. The manner of treating a Chinese character as a composition of radicals rather than a single character class largely reduces the size of vocabulary and enables RAN to possess the ability of recognizing unseen Chinese character classes, namely zero-shot learning.

Junbei Zhang, Xiaodan Zhu, Qian Chen et al.

The last several years have seen intensive interest in exploring neural-network-based models for machine comprehension (MC) and question answering (QA). In this paper, we approach the problems by closely modelling questions in a neural network framework. We first introduce syntactic information to help encode questions. We then view and model different types of questions and the information shared among them as an adaptation task and proposed adaptation models for them. On the Stanford Question Answering Dataset (SQuAD), we show that these approaches can help attain better results over a competitive baseline.

Shiliang Zhang, Cong Liu, Hui Jiang et al.

In this paper, we propose a novel neural network structure, namely \emph{feedforward sequential memory networks (FSMN)}, to model long-term dependency in time series without using recurrent feedback. The proposed FSMN is a standard fully-connected feedforward neural network equipped with some learnable memory blocks in its hidden layers. The memory blocks use a tapped-delay line structure to encode the long context information into a fixed-size representation as short-term memory mechanism. We have evaluated the proposed FSMNs in several standard benchmark tasks, including speech recognition and language modelling. Experimental results have shown FSMNs significantly outperform the conventional recurrent neural networks (RNN), including LSTMs, in modeling sequential signals like speech or language. Moreover, FSMNs can be learned much more reliably and faster than RNNs or LSTMs due to the inherent non-recurrent model structure.

ShiLiang Zhang, Hui Jiang, Si Wei et al.

We introduce a new structure for memory neural networks, called feedforward sequential memory networks (FSMN), which can learn long-term dependency without using recurrent feedback. The proposed FSMN is a standard feedforward neural networks equipped with learnable sequential memory blocks in the hidden layers. In this work, we have applied FSMN to several language modeling (LM) tasks. Experimental results have shown that the memory blocks in FSMN can learn effective representations of long history. Experiments have shown that FSMN based language models can significantly outperform not only feedforward neural network (FNN) based LMs but also the popular recurrent neural network (RNN) LMs.

Shiliang Zhang, Hui Jiang, Mingbin Xu et al.

In this paper, we propose the new fixed-size ordinally-forgetting encoding (FOFE) method, which can almost uniquely encode any variable-length sequence of words into a fixed-size representation. FOFE can model the word order in a sequence using a simple ordinally-forgetting mechanism according to the positions of words. In this work, we have applied FOFE to feedforward neural network language models (FNN-LMs). Experimental results have shown that without using any recurrent feedbacks, FOFE based FNN-LMs can significantly outperform not only the standard fixed-input FNN-LMs but also the popular RNN-LMs.