Search

Since the launch of ChatGPT at the end of 2022, generative dialogue models represented by ChatGPT have quickly become essential tools in daily life. As user expectations increase, enhancing the capability of generative dialogue models to solve complex problems has become a focal point of current research. This paper delves into the effectiveness of the RAFT (Retrieval Augmented Fine-Tuning) method in improving the performance of Generative dialogue models. RAFT combines chain-of-thought with model supervised fine-tuning (SFT) and retrieval augmented generation (RAG), which significantly enhanced the model's information extraction and logical reasoning abilities. We evaluated the RAFT method across multiple datasets and analysed its performance in various reasoning tasks, including long-form QA and short-form QA tasks, tasks in both Chinese and English, and supportive and comparison reasoning tasks. Notably, it addresses the gaps in previous research regarding long-form QA tasks and Chinese datasets. Moreover, we also evaluate the benefit of the chain-of-thought (CoT) in the RAFT method. This work offers valuable insights for studies focused on enhancing the performance of generative dialogue models., Comment: Accepted by ISCSLP 2024

The mainstream automatic speech recognition (ASR) technology usually requires hundreds to thousands of hours of annotated speech data. Three approaches to low-resourced ASR are phoneme or subword based supervised pre-training, and self-supervised pre-training over multilingual data. The Iu Mien language is the main ethnic language of the Yao ethnic group in China and is low-resourced in the sense that the annotated speech is very limited. With less than 10 hours of transcribed Iu Mien language, this paper investigates and compares the three approaches for Iu Mien speech recognition. Our experiments are based on the recently released, three backbone models pretrained over the 10 languages from the CommonVoice dataset (CV-Lang10), which correspond to the three approaches for low-resourced ASR. It is found that phoneme supervision can achieve better results compared to subword supervision and self-supervision, thereby providing higher data-efficiency. Particularly, the Whistle models, i.e., obtained by the weakly-supervised phoneme-based multilingual pre-training, obtain the most competitive results., Comment: Accepted into ISCSLP 2024

Streaming automatic speech recognition (ASR) is very important for many real-world ASR applications. However, a notable challenge for streaming ASR systems lies in balancing operational performance against latency constraint. Recently, a method of chunking, simulating future context and decoding, called CUSIDE, has been proposed for connectionist temporal classification (CTC) based streaming ASR, which obtains a good balance between reduced latency and high recognition accuracy. In this paper, we present CUSIDE-T, which successfully adapts the CUSIDE method over the recurrent neural network transducer (RNN-T) ASR architecture, instead of being based on the CTC architecture. We also incorporate language model rescoring in CUSIDE-T to further enhance accuracy, while only bringing a small additional latency. Extensive experiments are conducted over the AISHELL-1, WenetSpeech and SpeechIO datasets, comparing CUSIDE-T and U2++ (both based on RNN-T). U2++ is an existing counterpart of chunk based streaming ASR method. It is shown that CUSIDE-T achieves superior accuracy performance for streaming ASR, with equal settings of latency., Comment: Accepted into ISCSLP 2024

Recently multi-channel end-to-end (ME2E) ASR systems have emerged. While streaming single-channel end-to-end ASR has been extensively studied, streaming ME2E ASR is limited in exploration. Additionally, recent studies call attention to the gap between in-distribution (ID) and out-of-distribution (OOD) tests and doing realistic evaluations. This paper focuses on two research problems: realizing streaming ME2E ASR and improving OOD generalization. We propose the CUSIDE-array method, which integrates the recent CUSIDE methodology (Chunking, Simulating Future Context and Decoding) into the neural beamformer approach of ME2E ASR. It enables streaming processing of both front-end and back-end with a total latency of 402ms. The CUSIDE-array ME2E models are shown to achieve superior streaming results in both ID and OOD tests. Realistic evaluations confirm the advantage of CUSIDE-array in its capability to consume single-channel data to improve OOD generalization via back-end pre-training and ME2E fine-tuning., Comment: Accepted into ISCSLP 2024

There exist three approaches for multilingual and crosslingual automatic speech recognition (MCL-ASR) - supervised pre-training with phonetic or graphemic transcription, and self-supervised pre-training. We find that pre-training with phonetic supervision has been underappreciated so far for MCL-ASR, while conceptually it is more advantageous for information sharing between different languages. This paper explores the approach of pre-training with weakly phonetic supervision towards data-efficient MCL-ASR, which is called Whistle. We relax the requirement of gold-standard human-validated phonetic transcripts, and obtain International Phonetic Alphabet (IPA) based transcription by leveraging the LanguageNet grapheme-to-phoneme (G2P) models. We construct a common experimental setup based on the CommonVoice dataset, called CV-Lang10, with 10 seen languages and 2 unseen languages. A set of experiments are conducted on CV-Lang10 to compare, as fair as possible, the three approaches under the common setup for MCL-ASR. Experiments demonstrate the advantages of phoneme-based models (Whistle) for MCL-ASR, in terms of speech recognition for seen languages, crosslingual performance for unseen languages with different amounts of few-shot data, overcoming catastrophic forgetting, and training efficiency.It is found that when training data is more limited, phoneme supervision can achieve better results compared to subword supervision and self-supervision, thereby providing higher data-efficiency. To support reproducibility and promote future research along this direction, we will release the code, models and data for the whole pipeline of Whistle at https://github.com/thu-spmi/CAT upon publication.

Recently, increasing research interests have focused on retrieval augmented generation (RAG) to mitigate hallucination for large language models (LLMs). Following this trend, we launch the FutureDial-RAG challenge at SLT 2024, which aims at promoting the study of RAG for dialog systems. The challenge builds upon the MobileCS2 dataset, a real-life customer service datasets with nearly 3000 high-quality dialogs containing annotations for knowledge base query and corresponding results. Over the dataset, we define two tasks, track 1 for knowledge retrieval and track 2 for response generation, which are core research questions in dialog systems with RAG. We build baseline systems for the two tracks and design metrics to measure whether the systems can perform accurate retrieval and generate informative and coherent response. The baseline results show that it is very challenging to perform well on the two tasks, which encourages the participating teams and the community to study how to make better use of RAG for real-life dialog systems., Comment: Accepted by SLT 2024

Energy-Based Models (EBMs) are an important class of probabilistic models, also known as random fields and undirected graphical models. EBMs are un-normalized and thus radically different from other popular self-normalized probabilistic models such as hidden Markov models (HMMs), autoregressive models, generative adversarial nets (GANs) and variational auto-encoders (VAEs). Over the past years, EBMs have attracted increasing interest not only from the core machine learning community, but also from application domains such as speech, vision, natural language processing (NLP) and so on, due to significant theoretical and algorithmic progress. The sequential nature of speech and language also presents special challenges and needs a different treatment from processing fix-dimensional data (e.g., images). Therefore, the purpose of this monograph is to present a systematic introduction to energy-based models, including both algorithmic progress and applications in speech and language processing. First, the basics of EBMs are introduced, including classic models, recent models parameterized by neural networks, sampling methods, and various learning methods from the classic learning algorithms to the most advanced ones. Then, the application of EBMs in three different scenarios is presented, i.e., for modeling marginal, conditional and joint distributions, respectively. 1) EBMs for sequential data with applications in language modeling, where the main focus is on the marginal distribution of a sequence itself; 2) EBMs for modeling conditional distributions of target sequences given observation sequences, with applications in speech recognition, sequence labeling and text generation; 3) EBMs for modeling joint distributions of both sequences of observations and targets, and their applications in semi-supervised learning and calibrated natural language understanding., Comment: The version before publisher editing

Continual learning is crucial for dialog state tracking (DST) in dialog systems, since requirements from users for new functionalities are often encountered. However, most of existing continual learning methods for DST require task identities during testing, which is a severe limit in real-world applications. In this paper, we aim to address continual learning of DST in the class-incremental scenario (namely the task identity is unknown in testing). Inspired by the recently emerging prompt tuning method that performs well on dialog systems, we propose to use the prompt pool method, where we maintain a pool of key-value paired prompts and select prompts from the pool according to the distance between the dialog history and the prompt keys. The proposed method can automatically identify tasks and select appropriate prompts during testing. We conduct experiments on Schema-Guided Dialog dataset (SGD) and another dataset collected from a real-world dialog application. Experiment results show that the prompt pool method achieves much higher joint goal accuracy than the baseline. After combining with a rehearsal buffer, the model performance can be further improved.

Recent research has shown that multi-task pre-training greatly improves the model's robustness and transfer ability, which is crucial for building a high-quality dialog system. However, most previous works on multi-task pre-training rely heavily on human-defined input format or prompt, which is not optimal in quality and quantity. In this work, we propose to use Task-based Automatic Prompt generation (TAP) to automatically generate high-quality prompts. Using the high-quality prompts generated, we scale the corpus of the pre-trained conversation model to 122 datasets from 15 dialog-related tasks, resulting in Universal Pre-trained Conversation Model (UniPCM), a powerful foundation model for various conversational tasks and different dialog systems. Extensive experiments have shown that UniPCM is robust to input prompts and capable of various dialog-related tasks. Moreover, UniPCM has strong transfer ability and excels at low resource scenarios, achieving SOTA results on 9 different datasets ranging from task-oriented dialog to open-domain conversation. Furthermore, we are amazed to find that TAP can generate prompts on par with those collected with crowdsourcing. The code is released with the paper.

Most existing task-oriented dialog (TOD) systems track dialog states in terms of slots and values and use them to query a database to get relevant knowledge to generate responses. In real-life applications, user utterances are noisier, and thus it is more difficult to accurately track dialog states and correctly secure relevant knowledge. Recently, a progress in question answering and document-grounded dialog systems is retrieval-augmented methods with a knowledge retriever. Inspired by such progress, we propose a retrieval-based method to enhance knowledge selection in TOD systems, which significantly outperforms the traditional database query method for real-life dialogs. Further, we develop latent variable model based semi-supervised learning, which can work with the knowledge retriever to leverage both labeled and unlabeled dialog data. Joint Stochastic Approximation (JSA) algorithm is employed for semi-supervised model training, and the whole system is referred to as that JSA-KRTOD. Experiments are conducted on a real-life dataset from China Mobile Custom-Service, called MobileCS, and show that JSA-KRTOD achieves superior performances in both labeled-only and semi-supervised settings., Comment: 5 pages, accepted by INTERSPEECH2023

Energy-based language models (ELMs) parameterize an unnormalized distribution for natural sentences and are radically different from popular autoregressive language models (ALMs). As an important application, ELMs have been successfully used as a means for calculating sentence scores in speech recognition, but they all use less-modern CNN or LSTM networks. The recent progress in Transformer networks and large pretrained models such as BERT and GPT2 opens new possibility to further advancing ELMs. In this paper, we explore different architectures of energy functions and different training methods to investigate the capabilities of ELMs in rescoring for speech recognition, all using large pretrained models as backbones., Comment: Accepted into INTERSPEECH 2023

Maximum likelihood (ML) learning for energy-based models (EBMs) is challenging, partly due to non-convergence of Markov chain Monte Carlo.Several variations of ML learning have been proposed, but existing methods all fail to achieve both post-training image generation and proper density estimation. We propose to introduce diffusion data and learn a joint EBM, called diffusion assisted-EBMs, through persistent training (i.e., using persistent contrastive divergence) with an enhanced sampling algorithm to properly sample from complex, multimodal distributions. We present results from a 2D illustrative experiment and image experiments and demonstrate that, for the first time for image data, persistently trained EBMs can {\it simultaneously} achieve long-run stability, post-training image generation, and superior out-of-distribution detection., Comment: main text 8 pages

Building user simulators (USs) for reinforcement learning (RL) of task-oriented dialog systems (DSs) has gained more and more attention, which, however, still faces several fundamental challenges. First, it is unclear whether we can leverage pretrained language models to design, for example, GPT-2 based USs, to catch up and interact with the recently advanced GPT-2 based DSs. Second, an important ingredient in a US is that the user goal can be effectively incorporated and tracked; but how to flexibly integrate goal state tracking and develop an end-to-end trainable US for multi-domains has remained to be a challenge. In this work, we propose a generative user simulator (GUS) with GPT-2 based architecture and goal state tracking towards addressing the above two challenges. Extensive experiments are conducted on MultiWOZ2.1. Different DSs are trained via RL with GUS, the classic agenda-based user simulator (ABUS) and other ablation simulators respectively, and are compared for cross-model evaluation, corpus-based evaluation and human evaluation. The GUS achieves superior results in all three evaluation tasks., Comment: Accepted by EMNLP 2022 SereTOD Workshop

Recently, there has been progress in supervised funetuning pretrained GPT-2 to build end-to-end task-oriented dialog (TOD) systems. However, online reinforcement learning of a GPT-2 based dialog system (DS), together with a end-to-end user simulator (US), has not ever been explored. Moreover, a drawback with existing GPT-2 based TOD systems is that they mostly employ the whole dialog history as input, which brings inefficiencies in memory and compute. In this paper, we first propose Simplified Generative Architectures (SGA) for DS and US respectively, both based on GPT-2 but using shortened history. Then, we successfully develop Jointly Reinforced US and DS, called SGA-JRUD. Our DS with the proposed SGA, when only supervised trained, achieves state-of-the-art performance on MultiWOZ2.1 and is more compute-efficient in both training and generation. Extensive experiments on MultiWOZ2.1 further show the superiority of SGA-JRUD in both offline and online evaluations., Comment: An early version of Markovian Generative Architectures (MGA) and Generative User Simulator (GUS)

Recently, there have merged a class of task-oriented dialogue (TOD) datasets collected through Wizard-of-Oz simulated games. However, the Wizard-of-Oz data are in fact simulated data and thus are fundamentally different from real-life conversations, which are more noisy and casual. Recently, the SereTOD challenge is organized and releases the MobileCS dataset, which consists of real-world dialog transcripts between real users and customer-service staffs from China Mobile. Based on the MobileCS dataset, the SereTOD challenge has two tasks, not only evaluating the construction of the dialogue system itself, but also examining information extraction from dialog transcripts, which is crucial for building the knowledge base for TOD. This paper mainly presents a baseline study of the two tasks with the MobileCS dataset. We introduce how the two baselines are constructed, the problems encountered, and the results. We anticipate that the baselines can facilitate exciting future research to build human-robot dialogue systems for real-life tasks., Comment: Accepted by EMNLP 2022 SereTOD Workshop

Developing semi-supervised task-oriented dialog (TOD) systems by leveraging unlabeled dialog data has attracted increasing interests. For semi-supervised learning of latent state TOD models, variational learning is often used, but suffers from the annoying high-variance of the gradients propagated through discrete latent variables and the drawback of indirectly optimizing the target log-likelihood. Recently, an alternative algorithm, called joint stochastic approximation (JSA), has emerged for learning discrete latent variable models with impressive performances. In this paper, we propose to apply JSA to semi-supervised learning of the latent state TOD models, which is referred to as JSA-TOD. To our knowledge, JSA-TOD represents the first work in developing JSA based semi-supervised learning of discrete latent variable conditional models for such long sequential generation problems like in TOD systems. Extensive experiments show that JSA-TOD significantly outperforms its variational learning counterpart. Remarkably, semi-supervised JSA-TOD using 20% labels performs close to the full-supervised baseline on MultiWOZ2.1., Comment: Accepted into SIGDIAL 2022

A challenge on Semi-Supervised and Reinforced Task-Oriented Dialog Systems, Co-located with EMNLP2022 SereTOD Workshop., Comment: Version 2.1

Recently, Transformer based pretrained language models (PLMs), such as GPT2 and T5, have been leveraged to build generative task-oriented dialog (TOD) systems. A drawback of existing PLM-based models is their non-Markov architectures across turns, i.e., the whole history is used as the conditioning input at each turn. First, this brings inefficiencies in memory and computation. Furthermore, using the whole history increases model complexity and may hurt the training efficiency, especially when facing small amounts of labeled training data (the low-resource setting). In this paper, motivated by the observation that dialog states could be viewed as Markov states, we propose to build Markovian Generative Architectures (MGA) over PLM backbones for efficient TOD systems. Experiments on MultiWOZ2.1 show that in the rich-resource setting, the proposed Markov models reduce memory and time costs without performance degradation; in the low-resource setting, the training efficiency of the Markov models is more significant., Comment: Accepted by SLT 2022

Utilizing text-only data with an external language model (ELM) in end-to-end RNN-Transducer (RNN-T) for speech recognition is challenging. Recently, a class of methods such as density ratio (DR) and internal language model estimation (ILME) have been developed, outperforming the classic shallow fusion (SF) method. The basic idea behind these methods is that RNN-T posterior should first subtract the implicitly learned internal language model (ILM) prior, in order to integrate the ELM. While recent studies suggest that RNN-T only learns some low-order language model information, the DR method uses a well-trained neural language model with full context, which may be inappropriate for the estimation of ILM and deteriorate the integration performance. Based on the DR method, we propose a low-order density ratio method (LODR) by replacing the estimation with a low-order weak language model. Extensive empirical experiments are conducted on both in-domain and cross-domain scenarios on English LibriSpeech & Tedlium-2 and Chinese WenetSpeech & AISHELL-1 datasets. It is shown that LODR consistently outperforms SF in all tasks, while performing generally close to ILME and better than DR in most tests., Comment: Accepted into INTERSPEECH 2022

History and future contextual information are known to be important for accurate acoustic modeling. However, acquiring future context brings latency for streaming ASR. In this paper, we propose a new framework - Chunking, Simulating Future Context and Decoding (CUSIDE) for streaming speech recognition. A new simulation module is introduced to recursively simulate the future contextual frames, without waiting for future context. The simulation module is jointly trained with the ASR model using a self-supervised loss; the ASR model is optimized with the usual ASR loss, e.g., CTC-CRF as used in our experiments. Experiments show that, compared to using real future frames as right context, using simulated future context can drastically reduce latency while maintaining recognition accuracy. With CUSIDE, we obtain new state-of-the-art streaming ASR results on the AISHELL-1 dataset., Comment: Accepted into INTERSPEECH 2022

Recently, the end-to-end training approach for multi-channel ASR has shown its effectiveness, which usually consists of a beamforming front-end and a recognition back-end. However, the end-to-end training becomes more difficult due to the integration of multiple modules, particularly considering that multi-channel speech data recorded in real environments are limited in size. This raises the demand to exploit the single-channel data for multi-channel end-to-end ASR. In this paper, we systematically compare the performance of three schemes to exploit external single-channel data for multi-channel end-to-end ASR, namely back-end pre-training, data scheduling, and data simulation, under different settings such as the sizes of the single-channel data and the choices of the front-end. Extensive experiments on CHiME-4 and AISHELL-4 datasets demonstrate that while all three methods improve the multi-channel end-to-end speech recognition performance, data simulation outperforms the other two, at the cost of longer training time. Data scheduling outperforms back-end pre-training marginally but nearly consistently, presumably because that in the pre-training stage, the back-end tends to overfit on the single-channel data, especially when the single-channel data size is small., Comment: Accepted by ISCSLP 2022. arXiv admin note: substantial text overlap with arXiv:2107.02670

Recovering binary programs' call graphs is crucial for inter-procedural analysis tasks and applications based on them.transfer One of the core challenges is recognizing targets of indirect calls (i.e., indirect callees). Existing solutions all have high false positives and negatives, making call graphs inaccurate. In this paper, we propose a new solution Callee combining transfer learning and contrastive learning. The key insight is that, deep neural networks (DNNs) can automatically identify patterns concerning indirect calls, which can be more efficient than designing approximation algorithms or heuristic rules to handle various cases. Inspired by the advances in question-answering applications, we utilize contrastive learning to answer the callsite-callee question. However, one of the toughest challenges is that DNNs need large datasets to achieve high performance, while collecting large-scale indirect-call ground-truths can be computational-expensive. Since direct calls and indirect calls share similar calling conventions, it is possible to transfer knowledge learned from direct calls to indirect ones. Therefore, we leverage transfer learning to pre-train DNNs with easy-to-collect direct calls and further fine-tune the indirect-call DNNs. We evaluate Callee on several groups of targets, and results show that our solution could match callsites to callees with an F1-Measure of 94.6%, much better than state-of-the-art solutions. Further, we apply Callee to binary code similarity detection and hybrid fuzzing, and found it could greatly improve their performance.

Recently, two approaches, fine-tuning large pre-trained language models and variational training, have attracted significant interests, separately, for semi-supervised end-to-end task-oriented dialog (TOD) systems. In this paper, we propose Variational Latent-State GPT model (VLS-GPT), which is the first to combine the strengths of the two approaches. Among many options of models, we propose the generative model and the inference model for variational learning of the end-to-end TOD system, both as auto-regressive language models based on GPT-2, which can be further trained over a mix of labeled and unlabeled dialog data in a semi-supervised manner. Variational training of VLS-GPT is both statistically and computationally more challenging than previous variational learning works for sequential latent variable models, which use turn-level first-order Markovian. The inference model in VLS-GPT is non-Markovian due to the use of the Transformer architecture. In this work, we establish Recursive Monte Carlo Approximation (RMCA) to the variational objective with non-Markovian inference model and prove its unbiasedness. Further, we develop the computational strategy of sampling-then-forward-computation to realize RMCA, which successfully overcomes the memory explosion issue of using GPT in variational learning and speeds up training. Semi-supervised TOD experiments are conducted on two benchmark multi-domain datasets of different languages - MultiWOZ2.1 and CrossWOZ. VLS-GPT is shown to significantly outperform both supervised-only and semi-supervised self-training baselines., Comment: Accepted into IEEE/ACM Transactions on Audio, Speech and Language Processing

The use of phonological features (PFs) potentially allows language-specific phones to remain linked in training, which is highly desirable for information sharing for multilingual and crosslingual speech recognition methods for low-resourced languages. A drawback suffered by previous methods in using phonological features is that the acoustic-to-PF extraction in a bottom-up way is itself difficult. In this paper, we propose to join phonology driven phone embedding (top-down) and deep neural network (DNN) based acoustic feature extraction (bottom-up) to calculate phone probabilities. The new method is called JoinAP (Joining of Acoustics and Phonology). Remarkably, no inversion from acoustics to phonological features is required for speech recognition. For each phone in the IPA (International Phonetic Alphabet) table, we encode its phonological features to a phonological-vector, and then apply linear or nonlinear transformation of the phonological-vector to obtain the phone embedding. A series of multilingual and crosslingual (both zero-shot and few-shot) speech recognition experiments are conducted on the CommonVoice dataset (German, French, Spanish and Italian) and the AISHLL-1 dataset (Mandarin), and demonstrate the superiority of JoinAP with nonlinear phone embeddings over both JoinAP with linear phone embeddings and the traditional method with flat phone embeddings., Comment: ASRU2021

Automatic speech recognition systems have been largely improved in the past few decades and current systems are mainly hybrid-based and end-to-end-based. The recently proposed CTC-CRF framework inherits the data-efficiency of the hybrid approach and the simplicity of the end-to-end approach. In this paper, we further advance CTC-CRF based ASR technique with explorations on modeling units and neural architectures. Specifically, we investigate techniques to enable the recently developed wordpiece modeling units and Conformer neural networks to be succesfully applied in CTC-CRFs. Experiments are conducted on two English datasets (Switchboard, Librispeech) and a German dataset from CommonVoice. Experimental results suggest that (i) Conformer can improve the recognition performance significantly; (ii) Wordpiece-based systems perform slightly worse compared with phone-based systems for the target language with a low degree of grapheme-phoneme correspondence (e.g. English), while the two systems can perform equally strong when such degree of correspondence is high for the target language (e.g. German)., Comment: Submitted to ASRU 2021

Recently, the end-to-end training approach for neural beamformer-supported multi-channel ASR has shown its effectiveness in multi-channel speech recognition. However, the integration of multiple modules makes it more difficult to perform end-to-end training, particularly given that the multi-channel speech corpus recorded in real environments with a sizeable data scale is relatively limited. This paper explores the usage of single-channel data to improve the multi-channel end-to-end speech recognition system. Specifically, we design three schemes to exploit the single-channel data, namely pre-training, data scheduling, and data simulation. Extensive experiments on CHiME4 and AISHELL-4 datasets demonstrate that all three methods improve the multi-channel end-to-end training stability and speech recognition performance, while the data scheduling approach keeps a much simpler pipeline (vs. pre-training) and less computation cost (vs. data simulation). Moreover, we give a thorough analysis of our systems, including how the performance is affected by the choice of front-end, the data augmentation, training strategy, and single-channel data size., Comment: submitted to ASRU 2021

Time Delay Neural Networks (TDNNs) are widely used in both DNN-HMM based hybrid speech recognition systems and recent end-to-end systems. Nevertheless, the receptive fields of TDNNs are limited and fixed, which is not desirable for tasks like speech recognition, where the temporal dynamics of speech are varied and affected by many factors. This paper proposes to use deformable TDNNs for adaptive temporal dynamics modeling in end-to-end speech recognition. Inspired by deformable ConvNets, deformable TDNNs augment the temporal sampling locations with additional offsets and learn the offsets automatically based on the ASR criterion, without additional supervision. Experiments show that deformable TDNNs obtain state-of-the-art results on WSJ benchmarks (1.42\%/3.45\% WER on WSJ eval92/dev93 respectively), outperforming standard TDNNs significantly. Furthermore, we propose the latency control mechanism for deformable TDNNs, which enables deformable TDNNs to do streaming ASR without accuracy degradation., Comment: 5 pages. submitted to Interspeech 2021

Automatic speech recognition (ASR) has been significantly advanced with the use of deep learning and big data. However improving robustness, including achieving equally good performance on diverse speakers and accents, is still a challenging problem. In particular, the performance of children speech recognition (CSR) still lags behind due to 1) the speech and language characteristics of children's voice are substantially different from those of adults and 2) sizable open dataset for children speech is still not available in the research community. To address these problems, we launch the Children Speech Recognition Challenge (CSRC), as a flagship satellite event of IEEE SLT 2021 workshop. The challenge will release about 400 hours of Mandarin speech data for registered teams and set up two challenge tracks and provide a common testbed to benchmark the CSR performance. In this paper, we introduce the datasets, rules, evaluation method as well as baselines., Comment: 7 pages, 3 figures, 3 tables

Neural Architecture Search (NAS), the process of automating architecture engineering, is an appealing next step to advancing end-to-end Automatic Speech Recognition (ASR), replacing expert-designed networks with learned, task-specific architectures. In contrast to early computational-demanding NAS methods, recent gradient-based NAS methods, e.g., DARTS (Differentiable ARchiTecture Search), SNAS (Stochastic NAS) and ProxylessNAS, significantly improve the NAS efficiency. In this paper, we make two contributions. First, we rigorously develop an efficient NAS method via Straight-Through (ST) gradients, called ST-NAS. Basically, ST-NAS uses the loss from SNAS but uses ST to back-propagate gradients through discrete variables to optimize the loss, which is not revealed in ProxylessNAS. Using ST gradients to support sub-graph sampling is a core element to achieve efficient NAS beyond DARTS and SNAS. Second, we successfully apply ST-NAS to end-to-end ASR. Experiments over the widely benchmarked 80-hour WSJ and 300-hour Switchboard datasets show that the ST-NAS induced architectures significantly outperform the human-designed architecture across the two datasets. Strengths of ST-NAS such as architecture transferability and low computation cost in memory and time are also reported., Comment: Accepted by IEEE SLT 2021

A class of recent semi-supervised learning (SSL) methods heavily rely on domain-specific data augmentations. In contrast, generative SSL methods involve unsupervised learning based on generative models by either joint-training or pre-training, and are more appealing from the perspective of being domain-agnostic, since they do not inherently require data augmentations. Joint-training estimates the joint distribution of observations and labels, while pre-training is taken over observations only. Recently, energy-based models (EBMs) have achieved promising results for generative modeling. Joint-training via EBMs for SSL has been explored with encouraging results across different data modalities. In this paper, we make two contributions. First, we explore pre-training via EBMs for SSL and compare it to joint-training. Second, a suite of experiments are conducted over domains of image classification and natural language labeling to give a realistic whole picture of the performances of EBM based SSL methods. It is found that joint-training EBMs outperform pre-training EBMs marginally but nearly consistently.

Structured belief states are crucial for user goal tracking and database query in task-oriented dialog systems. However, training belief trackers often requires expensive turn-level annotations of every user utterance. In this paper we aim at alleviating the reliance on belief state labels in building end-to-end dialog systems, by leveraging unlabeled dialog data towards semi-supervised learning. We propose a probabilistic dialog model, called the LAtent BElief State (LABES) model, where belief states are represented as discrete latent variables and jointly modeled with system responses given user inputs. Such latent variable modeling enables us to develop semi-supervised learning under the principled variational learning framework. Furthermore, we introduce LABES-S2S, which is a copy-augmented Seq2Seq model instantiation of LABES. In supervised experiments, LABES-S2S obtains strong results on three benchmark datasets of different scales. In utilizing unlabeled dialog data, semi-supervised LABES-S2S significantly outperforms both supervised-only and semi-supervised baselines. Remarkably, we can reduce the annotation demands to 50% without performance loss on MultiWOZ., Comment: Accepted by EMNLP 2020

Although with progress in introducing auxiliary amortized inference models, learning discrete latent variable models is still challenging. In this paper, we show that the annoying difficulty of obtaining reliable stochastic gradients for the inference model and the drawback of indirectly optimizing the target log-likelihood can be gracefully addressed in a new method based on stochastic approximation (SA) theory of the Robbins-Monro type. Specifically, we propose to directly maximize the target log-likelihood and simultaneously minimize the inclusive divergence between the posterior and the inference model. The resulting learning algorithm is called joint SA (JSA). To the best of our knowledge, JSA represents the first method that couples an SA version of the EM (expectation-maximization) algorithm (SAEM) with an adaptive MCMC procedure. Experiments on several benchmark generative modeling and structured prediction tasks show that JSA consistently outperforms recent competitive algorithms, with faster convergence, better final likelihoods, and lower variance of gradient estimates., Comment: Accepted by UAI2020

In this paper, we present a new open source toolkit for speech recognition, named CAT (CTC-CRF based ASR Toolkit). CAT inherits the data-efficiency of the hybrid approach and the simplicity of the E2E approach, providing a full-fledged implementation of CTC-CRFs and complete training and testing scripts for a number of English and Chinese benchmarks. Experiments show CAT obtains state-of-the-art results, which are comparable to the fine-tuned hybrid models in Kaldi but with a much simpler training pipeline. Compared to existing non-modularized E2E models, CAT performs better on limited-scale datasets, demonstrating its data efficiency. Furthermore, we propose a new method called contextualized soft forgetting, which enables CAT to do streaming ASR without accuracy degradation. We hope CAT, especially the CTC-CRF based framework and software, will be of broad interest to the community, and can be further explored and improved., Comment: Accepted into INTERSPEECH 2020. arXiv admin note: text overlap with arXiv:1911.08747

Neural generative models have achieved promising performance on dialog generation tasks if given a huge data set. However, the lack of high-quality dialog data and the expensive data annotation process greatly limit their application in real-world settings. We propose a paraphrase augmented response generation (PARG) framework that jointly trains a paraphrase model and a response generation model to improve the dialog generation performance. We also design a method to automatically construct paraphrase training data set based on dialog state and dialog act labels. PARG is applicable to various dialog generation models, such as TSCP (Lei et al., 2018) and DAMD (Zhang et al., 2019). Experimental results show that the proposed framework improves these state-of-the-art dialog models further on CamRest676 and MultiWOZ. PARG also significantly outperforms other data augmentation methods in dialog generation tasks, especially under low resource settings., Comment: Accepted to ACL 2020, 10 pages, 2 figures

There has been a long recognition that discrete features (n-gram features) and neural network based features have complementary strengths for language models (LMs). Improved performance can be obtained by model interpolation, which is, however, a suboptimal two-step integration of discrete and neural features. The trans-dimensional random field (TRF) framework has the potential advantage of being able to flexibly integrate a richer set of features. However, either discrete or neural features are used alone in previous TRF LMs. This paper develops a mixed-feature TRF LM and demonstrates its advantage in integrating discrete and neural features. Various LMs are trained over PTB and Google one-billion-word datasets, and evaluated in N-best list rescoring experiments for speech recognition. Among all single LMs (i.e. without model interpolation), the mixed-feature TRF LMs perform the best, improving over both discrete TRF LMs and neural TRF LMs alone, and also being significantly better than LSTM LMs. Compared to interpolating two separately trained models with discrete and neural features respectively, the performance of mixed-feature TRF LMs matches the best interpolated model, and with simplified one-step training process and reduced training time., Comment: Accepted for presentation at ICASSP 2020, 5 pages, 2 figures

Conversations have an intrinsic one-to-many property, which means that multiple responses can be appropriate for the same dialog context. In task-oriented dialogs, this property leads to different valid dialog policies towards task completion. However, none of the existing task-oriented dialog generation approaches takes this property into account. We propose a Multi-Action Data Augmentation (MADA) framework to utilize the one-to-many property to generate diverse appropriate dialog responses. Specifically, we first use dialog states to summarize the dialog history, and then discover all possible mappings from every dialog state to its different valid system actions. During dialog system training, we enable the current dialog state to map to all valid system actions discovered in the previous process to create additional state-action pairs. By incorporating these additional pairs, the dialog policy learns a balanced action distribution, which further guides the dialog model to generate diverse responses. Experimental results show that the proposed framework consistently improves dialog policy diversity, and results in improved response diversity and appropriateness. Our model obtains state-of-the-art results on MultiWOZ.

In this paper, we present a new open source toolkit for automatic speech recognition (ASR), named CAT (CRF-based ASR Toolkit). A key feature of CAT is discriminative training in the framework of conditional random field (CRF), particularly with connectionist temporal classification (CTC) inspired state topology. CAT contains a full-fledged implementation of CTC-CRF and provides a complete workflow for CRF-based end-to-end speech recognition. Evaluation results on Chinese and English benchmarks such as Switchboard and Aishell show that CAT obtains the state-of-the-art results among existing end-to-end models with less parameters, and is competitive compared with the hybrid DNN-HMM models. Towards flexibility, we show that i-vector based speaker-adapted recognition and latency control mechanism can be explored easily and effectively in CAT. We hope CAT, especially the CRF-based framework and software, will be of broad interest to the community, and can be further explored and improved., Comment: Code released at: https://github.com/thu-spmi/cat

Conditional random fields (CRFs) have been shown to be one of the most successful approaches to sequence labeling. Various linear-chain neural CRFs (NCRFs) are developed to implement the non-linear node potentials in CRFs, but still keeping the linear-chain hidden structure. In this paper, we propose NCRF transducers, which consists of two RNNs, one extracting features from observations and the other capturing (theoretically infinite) long-range dependencies between labels. Different sequence labeling methods are evaluated over POS tagging, chunking and NER (English, Dutch). Experiment results show that NCRF transducers achieve consistent improvements over linear-chain NCRFs and RNN transducers across all the four tasks, and can improve state-of-the-art results.

Slot filling is a crucial component in task-oriented dialog systems that is used to parse (user) utterances into semantic concepts called slots. An ontology is defined by the collection of slots and the values that each slot can take. The most widely used practice of treating slot filling as a sequence labeling task suffers from two main drawbacks. First, the ontology is usually pre-defined and fixed and therefore is not able to detect new labels for unseen slots. Second, the one-hot encoding of slot labels ignores the correlations between slots with similar semantics, which makes it difficult to share knowledge learned across different domains. To address these problems, we propose a new model called elastic conditional random field (eCRF), where each slot is represented by the embedding of its natural language description and modeled by a CRF layer. New slot values can be detected by eCRF whenever a language description is available for the slot. In our experiment, we show that eCRFs outperform existing models in both in-domain and cross-domain tasks, especially in predicting unseen slots and values., Comment: 5 pages

This document aims to provide a review on learning with deep generative models (DGMs), which is an highly-active area in machine learning and more generally, artificial intelligence. This review is not meant to be a tutorial, but when necessary, we provide self-contained derivations for completeness. This review has two features. First, though there are different perspectives to classify DGMs, we choose to organize this review from the perspective of graphical modeling, because the learning methods for directed DGMs and undirected DGMs are fundamentally different. Second, we differentiate model definitions from model learning algorithms, since different learning algorithms can be applied to solve the learning problem on the same model, and an algorithm can be applied to learn different models. We thus separate model definition and model learning, with more emphasis on reviewing, differentiating and connecting different learning algorithms. We also discuss promising future research directions., Comment: add SN-GANs, SA-GANs, conditional generation (cGANs, AC-GANs). arXiv admin note: text overlap with arXiv:1606.00709, arXiv:1801.03558 by other authors

In this paper, we present an end-to-end automatic speech recognition system, which successfully employs subword units in a hybrid CTC-Attention based system. The subword units are obtained by the byte-pair encoding (BPE) compression algorithm. Compared to using words as modeling units, using characters or subword units does not suffer from the out-of-vocabulary (OOV) problem. Furthermore, using subword units further offers a capability in modeling longer context than using characters. We evaluate different systems over the LibriSpeech 1000h dataset. The subword-based hybrid CTC-Attention system obtains 6.8% word error rate (WER) on the test_clean subset without any dictionary or external language model. This represents a significant improvement (a 12.8% WER relative reduction) over the character-based hybrid CTC-Attention system., Comment: accepted by ISCSLP 2018

A new whole-sentence language model - neural trans-dimensional random field language model (neural TRF LM), where sentences are modeled as a collection of random fields, and the potential function is defined by a neural network, has been introduced and successfully trained by noise-contrastive estimation (NCE). In this paper, we extend NCE and propose dynamic noise-contrastive estimation (DNCE) to solve the two problems observed in NCE training. First, a dynamic noise distribution is introduced and trained simultaneously to converge to the data distribution. This helps to significantly cut down the noise sample number used in NCE and reduce the training cost. Second, DNCE discriminates between sentences generated from the noise distribution and sentences generated from the interpolation of the data distribution and the noise distribution. This alleviates the overfitting problem caused by the sparseness of the training set. With DNCE, we can successfully and efficiently train neural TRF LMs on large corpus (about 0.8 billion words) with large vocabulary (about 568 K words). Neural TRF LMs perform as good as LSTM LMs with less parameters and being 5x~114x faster in rescoring sentences. Interpolating neural TRF LMs with LSTM LMs and n-gram LMs can further reduce the error rates., Comment: 6 pages. Has been submitted to SLT 2018

End-to-end speaker verification systems have received increasing interests. The traditional i-vector approach trains a generative model (basically a factor-analysis model) to extract i-vectors as speaker embeddings. In contrast, the end-to-end approach directly trains a discriminative model (often a neural network) to learn discriminative speaker embeddings; a crucial component is the training criterion. In this paper, we use angular softmax (A-softmax), which is originally proposed for face verification, as the loss function for feature learning in end-to-end speaker verification. By introducing margins between classes into softmax loss, A-softmax can learn more discriminative features than softmax loss and triplet loss, and at the same time, is easy and stable for usage. We make two contributions in this work. 1) We introduce A-softmax loss into end-to-end speaker verification and achieve significant EER reductions. 2) We find that the combination of using A-softmax in training the front-end and using PLDA in the back-end scoring further boosts the performance of end-to-end systems under short utterance condition (short in both enrollment and test). Experiments are conducted on part of $Fisher$ dataset and demonstrate the improvements of using A-softmax.

Neural random fields (NRFs), referring to a class of generative models that use neural networks to implement potential functions in random fields (a.k.a. energy-based models), are not new but receive less attention with slow progress. Different from various directed graphical models such as generative adversarial networks (GANs), NRFs provide an interesting family of undirected graphical models for generative modeling. In this paper we propose a new approach, the inclusive-NRF approach, to learning NRFs for continuous data (e.g. images), by introducing inclusive-divergence minimized auxiliary generators and developing stochastic gradient sampling in an augmented space. Based on the new approach, specific inclusive-NRF models are developed and thoroughly evaluated in two important generative modeling applications - image generation and anomaly detection. The proposed models consistently improve over state-of-the-art results in both applications. Remarkably, in addition to superior sample generation, one additional benefit of our inclusive-NRF approach is that, unlike GANs, it can directly provide (unnormalized) density estimate for sample evaluation. With these contributions and results, this paper significantly advances the learning and applications of NRFs to a new level, both theoretically and empirically, which have never been obtained before.

An ensemble of neural networks is known to be more robust and accurate than an individual network, however usually with linearly-increased cost in both training and testing. In this work, we propose a two-stage method to learn Sparse Structured Ensembles (SSEs) for neural networks. In the first stage, we run SG-MCMC with group sparse priors to draw an ensemble of samples from the posterior distribution of network parameters. In the second stage, we apply weight-pruning to each sampled network and then perform retraining over the remained connections. In this way of learning SSEs with SG-MCMC and pruning, we not only achieve high prediction accuracy since SG-MCMC enhances exploration of the model-parameter space, but also reduce memory and computation cost significantly in both training and testing of NN ensembles. This is thoroughly evaluated in the experiments of learning SSE ensembles of both FNNs and LSTMs. For example, in LSTM based language modeling (LM), we obtain 21% relative reduction in LM perplexity by learning a SSE of 4 large LSTM models, which has only 30% of model parameters and 70% of computations in total, as compared to the baseline large LSTM LM. To the best of our knowledge, this work represents the first methodology and empirical study of integrating SG-MCMC, group sparse prior and network pruning together for learning NN ensembles.

Dialog state tracking (DST) is a crucial component in a task-oriented dialog system for conversational information access. A common practice in current dialog systems is to define the dialog state by a set of slot-value pairs. Such representation of dialog states and the slot-filling based DST have been widely employed, but suffer from three drawbacks. (1) The dialog state can contain only a single value for a slot, and (2) can contain only users' affirmative preference over the values for a slot. (3) Current task-based dialog systems mainly focus on the searching task, while the enquiring task is also very common in practice. The above observations motivate us to enrich current representation of dialog states and collect a brand new dialog dataset about movies, based upon which we build a new DST, called enriched DST (EDST), for flexible accessing movie information. The EDST supports the searching task, the enquiring task and their mixed task. We show that the new EDST method not only achieves good results on Iqiyi dataset, but also outperforms other state-of-the-art DST methods on the traditional dialog datasets, WOZ2.0 and DSTC2., Comment: 5 pages, 2 figures, accepted by ICASSP2018

Trans-dimensional random field language models (TRF LMs) where sentences are modeled as a collection of random fields, have shown close performance with LSTM LMs in speech recognition and are computationally more efficient in inference. However, the training efficiency of neural TRF LMs is not satisfactory, which limits the scalability of TRF LMs on large training corpus. In this paper, several techniques on both model formulation and parameter estimation are proposed to improve the training efficiency and the performance of neural TRF LMs. First, TRFs are reformulated in the form of exponential tilting of a reference distribution. Second, noise-contrastive estimation (NCE) is introduced to jointly estimate the model parameters and normalization constants. Third, we extend the neural TRF LMs by marrying the deep convolutional neural network (CNN) and the bidirectional LSTM into the potential function to extract the deep hierarchical features and bidirectionally sequential features. Utilizing all the above techniques enables the successful and efficient training of neural TRF LMs on a 40x larger training set with only 1/3 training time and further reduces the WER with relative reduction of 4.7% on top of a strong LSTM LM baseline., Comment: 5 pages and 2 figures

Trans-dimensional random field language models (TRF LMs) have recently been introduced, where sentences are modeled as a collection of random fields. The TRF approach has been shown to have the advantages of being computationally more efficient in inference than LSTM LMs with close performance and being able to flexibly integrating rich features. In this paper we propose neural TRFs, beyond of the previous discrete TRFs that only use linear potentials with discrete features. The idea is to use nonlinear potentials with continuous features, implemented by neural networks (NNs), in the TRF framework. Neural TRFs combine the advantages of both NNs and TRFs. The benefits of word embedding, nonlinear feature learning and larger context modeling are inherited from the use of NNs. At the same time, the strength of efficient inference by avoiding expensive softmax is preserved. A number of technical contributions, including employing deep convolutional neural networks (CNNs) to define the potentials and incorporating the joint stochastic approximation (JSA) strategy in the training algorithm, are developed in this work, which enable us to successfully train neural TRF LMs. Various LMs are evaluated in terms of speech recognition WERs by rescoring the 1000-best lists of WSJ'92 test data. The results show that neural TRF LMs not only improve over discrete TRF LMs, but also perform slightly better than LSTM LMs with only one fifth of parameters and 16x faster inference efficiency., Comment: 6 pages, 2 figures and 3 tables, accepted to ASRU 2017

State-of-the-art i-vector based speaker verification relies on variants of Probabilistic Linear Discriminant Analysis (PLDA) for discriminant analysis. We are mainly motivated by the recent work of the joint Bayesian (JB) method, which is originally proposed for discriminant analysis in face verification. We apply JB to speaker verification and make three contributions beyond the original JB. 1) In contrast to the EM iterations with approximated statistics in the original JB, the EM iterations with exact statistics are employed and give better performance. 2) We propose to do simultaneous diagonalization (SD) of the within-class and between-class covariance matrices to achieve efficient testing, which has broader application scope than the SVD-based efficient testing method in the original JB. 3) We scrutinize similarities and differences between various Gaussian PLDAs and JB, complementing the previous analysis of comparing JB only with Prince-Elder PLDA. Extensive experiments are conducted on NIST SRE10 core condition 5, empirically validating the superiority of JB with faster convergence rate and 9-13% EER reduction compared with state-of-the-art PLDA., Comment: accepted by ICASSP2017

The dominant language models (LMs) such as n-gram and neural network (NN) models represent sentence probabilities in terms of conditionals. In contrast, a new trans-dimensional random field (TRF) LM has been recently introduced to show superior performances, where the whole sentence is modeled as a random field. In this paper, we examine how the TRF models can be interpolated with the NN models, and obtain 12.1\% and 17.9\% relative error rate reductions over 6-gram LMs for English and Chinese speech recognition respectively through log-linear combination., Comment: three pages, 2 experiment result tables, reporting the WERs on an Englisth dateset and a Chinese dataset