We introduce Florence-2, a novel vision foundation model with a unified, prompt-based representation for various computer vision and vision-language tasks. While existing large vision models excel in transfer learning, they struggle to perform diverse tasks with simple instructions, a capability that implies handling the complexity of various spatial hierarchy and semantic granularity. Florence-2 was designed to take text-prompt as task instructions and generate desirable results in text forms, whether it be captioning, object detection, grounding or segmentation. This multi-task learning setup demands large-scale, high-quality annotated data. To this end, we co-developed FLD-5B that consists of 5.4 billion comprehensive visual annotations on 126 million images, using an iterative strategy of automated image annotation and model refinement. We adopted a sequence-to-sequence structure to train Florence-2 to perform versatile and comprehensive vision tasks. Extensive evaluations on numerous tasks demonstrated Florence-2 to be a strong vision foundation model contender with un-precedented zero-shot and fine-tuning capabilities.

Most end-to-end (E2E) speech recognition models are composed of encoder and decoder blocks that perform acoustic and language modeling functions. Pretrained large language models (LLMs) have the potential to improve the performance of E2E ASR. However, integrating a pretrained language model into an E2E speech recognition model has shown limited benefits due to the mismatches between text-based LLMs and those used in E2E ASR. In this paper, we explore an alternative approach by adapting a pretrained LLMs to speech. Our experiments on fully-formatted E2E ASR transcription tasks across various domains demonstrate that our approach can effectively leverage the strengths of pretrained LLMs to produce more readable ASR transcriptions. Our model, which is based on the pretrained large language models with either an encoder-decoder or decoder-only structure, surpasses strong ASR models such as Whisper1, in terms of recognition error rate, considering formats like punctuation and capitalization as well.

There is a rising interest and trend in research towards directly translating speech from one language to another, known as end-to-end speech-to-speech translation. However, most end-to-end models struggle to outperform cascade models, i.e., a pipeline framework by concatenating speech recognition, machine translation and text-to-speech models. The primary challenges stem from the inherent complexities involved in direct translation tasks and the scarcity of data. In this study, we introduce a novel model framework TransVIP that leverages diverse datasets in a cascade fashion yet facilitates end-to-end inference through joint probability. Furthermore, we propose two separated encoders to preserve the speaker’s voice characteristics and isochrony from the source speech during the translation process, making it highly suitable for scenarios such as video dubbing. Our experiments on the French-English language pair demonstrate that our model outperforms the current state-of-the-art speech-to-speech translation model.

Recent advancements in zero-shot text-to-speech (TTS) modeling have led to significant strides in generating high-fidelity and diverse speech. However, dialogue generation, along with achieving human-like naturalness in speech, continues to be a challenge. In this paper, we introduce CoVoMix: Conversational Voice Mixture Generation, a novel model for zero-shot, human-like, multi-speaker, multi-round dialogue speech generation. CoVoMix first converts dialogue text into multiple streams of discrete tokens, with each token stream representing semantic information for individual talkers. These token streams are then fed into a flow-matching based acoustic model to generate mixed mel-spectrograms. Finally, the speech waveforms are produced using a HiFi-GAN model. Furthermore, we devise a comprehensive set of metrics for measuring the effectiveness of dialogue modeling and generation. Our experimental results show that CoVoMix can generate dialogues that are not only human-like in their naturalness and coherence but also involve multiple talkers engaging in multiple rounds of conversation. This is exemplified by instances generated in a single channel where one speaker's utterance is seamlessly mixed with another's interjections or laughter, indicating the latter's role as an attentive listener. Audio samples are enclosed in the supplementary.

Neural audio codec tokens serve as the fundamental building blocks for speech language model (SLM)-based speech generation. However, there is no systematic understanding on how the codec system affects the speech generation performance of the SLM. In this work, we examine codec tokens within SLM framework for speech generation to provide insights for effective codec design. We retrain existing high-performing neural codec models on the same data set and loss functions to compare their performance in a uniform setting. We integrate codec tokens into two SLM systems: masked-based parallel speech generation system and an auto-regressive (AR) plus non-auto-regressive (NAR) model-based system. Our findings indicate that better speech reconstruction in codec systems does not guarantee improved speech generation in SLM. A high-quality codec decoder is crucial for natural speech production in SLM, while speech intelligibility depends more on quantization mechanism.

Human intelligence is multimodal; we integrate visual, linguistic, and acoustic signals to maintain a holistic worldview. Most current pretraining methods, however, are limited to one or two modalities. We present i-Code, a self-supervised pretraining framework where users may flexibly combine the modalities of vision, speech, and language into unified and general-purpose vector representations. In this framework, data from each modality are first given to pretrained single-modality encoders. The encoder outputs are then integrated with a multimodal fusion network, which uses novel merge- and co-attention mechanisms to effectively combine information from the different modalities. The entire system is pretrained end-to-end with new objectives including masked modality unit modeling and cross-modality contrastive learning. Unlike previous research using only video for pretraining, the i-Code framework can dynamically process single, dual, and triple-modality data during training and inference, flexibly projecting different combinations of modalities into a single representation space. Experimental results demonstrate how i-Code can outperform state-of-the-art techniques on five multimodal understanding tasks and single-modality benchmarks, improving by as much as 11% and demonstrating the power of integrative multimodal pretraining.

Yulong Chen, Yang Liu, Ruochen Xu, Ziyi Yang, Chenguang Zhu, Michael Zeng, Yue Zhang. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Pengcheng He, Baolin Peng, Song Wang, Yang Liu, Ruochen Xu, Hany Hassan, Yu Shi, Chenguang Zhu, Wayne Xiong, Michael Zeng, Jianfeng Gao, Xuedong Huang. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

We propose Universal Document Processing (UDOP), a foundation Document AI model which unifies text, image, and layout modalities together with varied task formats, including document understanding and generation. UDOP leverages the spatial correlation between textual content and document image to model image, text, and layout modalities with one uniform representation. With a novel Vision-Text-Layout Transformer, UDOP unifies pretraining and multi-domain downstream tasks into a prompt-based sequence generation scheme. UDOP is pretrained on both large-scale unlabeled document corpora using innovative self-supervised objectives and diverse labeled data. UDOP also learns to generate document images from text and layout modalities via masked image reconstruction. To the best of our knowledge, this is the first time in the field of document AI that one model simultaneously achieves high-quality neural document editing and content customization. Our method sets the state-of-the-art on 8 Document AI tasks, e.g., document understanding and QA, across diverse data domains like finance reports, academic papers, and web-sites. UDOP ranks first on the leaderboard of the Document Understanding Benchmark.11Code and models: https://github.com/microsoft/i-Code/tree/main/i-Code-Doc

Recently, large-scale text-to-image (T2I) models have shown impressive performance in generating high-fidelity images, but with limited controllability, e.g., precisely specifying the content in a specific region with a free-form text description. In this paper, we propose an effective technique for such regional control in T2I generation. We augment T2I models' inputs with an extra set of position tokens, which represent the quantized spatial coordinates. Each region is specified by four position tokens to represent the top-left and bottom-right corners, followed by an open-ended natural language regional description. Then, we fine-tune a pre-trained T2I model with such new input interface. Our model, dubbed as ReCo (Region-Controlled T2I), enables the region control for arbitrary objects described by open-ended regional texts rather than by object labels from a constrained category set. Empirically, ReCo achieves better image quality than the T2I model strengthened by positional words (FID: 8.82 → 7.36, SceneFID: 15.54 → 6.51 on COCO), together with objects being more accurately placed, amounting to a 20.40% region classification accuracy improvement on COCO. Furthermore, we demonstrate that ReCo can better control the object count, spatial relationship, and region attributes such as color/size, with the free-form regional description. Human evaluation on PaintSkill shows that ReCo is +19.28% and +17.21% more accurate in generating images with correct object count and spatial relationship than the T2I model. Code is available at https://github.com/microsoft/Reeo.

Large Language Models (LLMs) have shown impressive performance as general purpose agents, but their abilities remain highly dependent on prompts which are hand written with onerous trial-and-error effort.We propose a simple and nonparametric solution to this problem, Prompt Optimization with Textual Gradients (ProTeGi), which is inspired by numerical gradient descent to automatically improve prompts, assuming access to training data and an LLM API.The algorithm uses minibatches of data to form natural language "gradients" that criticize the current prompt, much like how numerical gradients point in the direction of error ascent.The natural language gradients are then "propagated" into the prompt by editing the prompt in the opposite semantic direction of the gradient.These gradient descent steps are guided by a beam search and bandit selection procedure which significantly improves algorithmic efficiency.Preliminary results across three benchmark NLP tasks and the novel problem of LLM jailbreak detection suggest that Automatic Prompt Optimization can outperform prior prompt editing techniques and improve an initial prompt's performance by up to 31%, by using data to rewrite vague task descriptions into more precise annotation instructions.1

Automatic target sound extraction (TSE) is a machine learning approach to mimic the human auditory perception capability of attending to a sound source of interest from a mixture of sources. It often uses a model conditioned on a fixed form of target sound clues, such as a sound class label, which limits the ways in which users can interact with the model to specify the target sounds. To leverage variable number of clues cross modalities available in the inference phase, including a video, a sound event class, and a text caption, we propose a unified transformer-based TSE model architecture, where a multi-clue attention module integrates all the clues across the modalities. Since there is no off-the-shelf benchmark to evaluate our proposed approach, we build a dataset1based on public corpora, Audioset and AudioCaps. Experimental results for seen and unseen target-sound evaluation sets show that our proposed TSE model can effectively deal with a varying number of clues which improves the TSE performance and robustness against partially compromised clues.

Self-supervised learning has been successfully applied to various speech recognition and understanding tasks. However, for generative tasks such as speech enhancement and speech separation, most self-supervised speech representations did not show substantial improvements. To deal with this problem, in this paper, we propose data2vec-SG (Speech Generation), which is a teacher-student learning framework that addresses speech generation tasks. Our data2vec-SG introduces a reconstruction module into data2vec [1] and enforces the representations to contain not only the semantic information but also the acoustic knowledge to generate clean speech waveforms. Experimental results demonstrate that the proposed framework boosts the performance of various speech generation tasks including speech enhancement, speech separation, and packet loss concealment. Meanwhile, the learned representation is also capable of helping other downstream tasks, which is demonstrated by the good performance in the speech recognition task in both clean and noisy conditions.

Code-switching speech refers to a means of expression by mixing two or more languages within a single utterance. Automatic Speech Recognition (ASR) with End-to-End (E2E) modeling for such speech can be a challenging task due to the lack of data. In this study, we investigate text generation and injection for improving the performance of an industry commonly-used streaming model, Transformer-Transducer (T-T), in Mandarin-English code-switching speech recognition. We first propose a strategy to generate codeswitching text data and then investigate injecting generated text into T-T model explicitly by Text-To-Speech (TTS) conversion or implicitly by tying speech and text latent spaces. Experimental results on the T-T model trained with a dataset containing 1,800 hours of real Mandarin-English code-switched speech show that our approaches to inject generated code-switching text significantly boost the performance of T-T models, i.e., 16% relative Token-based Error Rate (TER) reduction averaged on three evaluation sets, and the approach of tying speech and text latent spaces is superior to that of TTS conversion on the evaluation set which contains more homogeneous data with the training set.

Joint speech-language training is challenging due to the large demand for training data and GPU consumption, as well as the modality gap between speech and language. We present ComSL, a speech-language model built atop a composite architecture of public pre-trained speech-only and language-only models and optimized data-efficiently for spoken language tasks. Particularly, we propose to incorporate cross-modality learning into transfer learning and conduct them simultaneously for downstream tasks in a multi-task learning manner. Our approach has demonstrated effectiveness in end-to-end speech-to-text translation tasks, achieving a new state-of-the-art average BLEU score of 31.5 on the multilingual speech to English text translation task for 21 languages, as measured on the public CoVoST2 evaluation set.

We present Composable Diffusion (CoDi), a novel generative model capable of generating any combination of output modalities, such as language, image, video, or audio, from any combination of input modalities. Unlike existing generative AI systems, CoDi can generate multiple modalities in parallel and its input is not limited to a subset of modalities like text or image. Despite the absence of training datasets for many combinations of modalities, we propose to align modalities in both the input and output space. This allows CoDi to freely condition on any input combination and generate any group of modalities, even if they are not present in the training data. CoDi employs a novel composable generation strategy which involves building a shared multimodal space by bridging alignment in the diffusion process, enabling the synchronized generation of intertwined modalities, such as temporally aligned video and audio. Highly customizable and flexible, CoDi achieves strong joint-modality generation quality, and outperforms or is on par with the unimodal state-of-the-art for single-modality synthesis.

Abstract Controllable summarization allows users to generate customized summaries with specified attributes. However, due to the lack of designated annotations of controlled summaries, existing work has to craft pseudo datasets by adapting generic summarization benchmarks. Furthermore, most research focuses on controlling single attributes individually (e.g., a short summary or a highly abstractive summary) rather than controlling a mix of attributes together (e.g., a short and highly abstractive summary). In this paper, we propose MACSum, the first human-annotated summarization dataset for controlling mixed attributes. It contains source texts from two domains, news articles and dialogues, with human-annotated summaries controlled by five designed attributes (Length, Extractiveness, Specificity, Topic, and Speaker). We propose two simple and effective parameter-efficient approaches for the new task of mixed controllable summarization based on hard prompt tuning and soft prefix tuning. Results and analysis demonstrate that hard prompt models yield the best performance on most metrics and human evaluations. However, mixed-attribute control is still challenging for summarization tasks. Our dataset and code are available at https://github.com/psunlpgroup/MACSum.

Modern Automatic Speech Recognition (ASR) systems can achieve high performance in terms of recognition accuracy. However, a perfectly accurate transcript still can be challenging to read due to grammatical errors, disfluency, and other noises common in spoken communication. These readable issues introduced by speakers and ASR systems will impair the performance of downstream tasks and the understanding of human readers. In this work, we present a task called ASR post-processing for readability (APR) and formulate it as a sequence-to-sequence text generation problem. The APR task aims to transform the noisy ASR output into a readable text for humans and downstream tasks while maintaining the semantic meaning of speakers. We further study the APR task from the benchmark dataset, evaluation metrics, and baseline models: First, to address the lack of task-specific data, we propose a method to construct a dataset for the APR task by using the data collected for grammatical error correction. Second, we utilize metrics adapted or borrowed from similar tasks to evaluate model performance on the APR task. Lastly, we use several typical or adapted pre-trained models as the baseline models for the APR task. Furthermore, we fine-tune the baseline models on the constructed dataset and compare their performance with a traditional pipeline method in terms of proposed evaluation metrics. Experimental results show that all the fine-tuned baseline models perform better than the traditional pipeline method, and our adapted RoBERTa model outperforms the pipeline method by 4.95 and 6.63 BLEU points on two test sets, respectively. The human evaluation and case study further reveal the ability of the proposed model to improve the readability of ASR transcripts.

Knowledge graphs (KGs) contain rich information about world knowledge, entities, and relations. Thus, they can be great supplements to existing pre-trained language models. However, it remains a challenge to efficiently integrate information from KG into language modeling. And the understanding of a knowledge graph requires related context. We propose a novel joint pre-training framework, JAKET, to model both the knowledge graph and language. The knowledge module and language module provide essential information to mutually assist each other: the knowledge module produces embeddings for entities in text while the language module generates context-aware initial embeddings for entities and relations in the graph. Our design enables the pre-trained model to easily adapt to unseen knowledge graphs in new domains. Experiment results on several knowledge-aware NLP tasks show that our proposed framework achieves superior performance by effectively leveraging knowledge in language understanding.

Dialogue is an essential part of human communication and cooperation. Existing research mainly focuses on short dialogue scenarios in a one-on-one fashion. However, multi-person interactions in the real world, such as meetings or interviews, are frequently over a few thousand words. There is still a lack of corresponding research and powerful tools to understand and process such long dialogues. Therefore, in this work, we present a pre-training framework for long dialogue understanding and summarization. Considering the nature of long conversations, we propose a window-based denoising approach for generative pre-training. For a dialogue, it corrupts a window of text with dialogue-inspired noise, and guides the model to reconstruct this window based on the content of the remaining conversation. Furthermore, to process longer input, we augment the model with sparse attention which is combined with conventional attention in a hybrid manner. We conduct extensive experiments on five datasets of long dialogues, covering tasks of dialogue summarization, abstractive question answering and topic segmentation. Experimentally, we show that our pre-trained model DialogLM significantly surpasses the state-of-the-art models across datasets and tasks. Source code and all the pre-trained models are available on our GitHub repository (https://github.com/microsoft/DialogLM).

Shuohang Wang, Yichong Xu, Yuwei Fang, Yang Liu, Siqi Sun, Ruochen Xu, Chenguang Zhu, Michael Zeng. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Donghan Yu, Chenguang Zhu, Yuwei Fang, Wenhao Yu, Shuohang Wang, Yichong Xu, Xiang Ren, Yiming Yang, Michael Zeng. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Vision-and-language (VL) pre-training has proven to be highly effective on various VL downstream tasks. While recent work has shown that fully transformer-based VL models can be more efficient than previous region-feature-based methods, their performance on downstream tasks often degrades significantly. In this paper, we present Meter, a Multimodal End-to-end TransformER framework, through which we investigate how to design and pre-train a fully transformer-based VL model in an end-to-end manner. Specifically, we dissect the model designs along multiple dimensions: vision encoders (e.g., CLIP-ViT, Swin transformer), text encoders (e.g., RoBERTa, De-BERTa), multimodal fusion module (e.g., merged attention vs. co-attention), architectural design (e.g., encoder-only vs. encoder-decoder), and pre-training objectives (e.g., masked image modeling). We conduct comprehensive experiments and provide insights on how to train a performant VL transformer. Meterachieves an accuracy of 77.64% on the VQAv2 test-std set using only 4M images for pre-training, surpassing the state-of-the-art region-feature-based model by 1.04%, and outperforming the previous best fully transformer-based model by 1.6%. Notably, when further scaled up, our best VQA model achieves an accuracy of 80.54%. Code and pre-trained models are released at https://github.com/zdou0830/METER.

Vision-language (V+L) pretraining models have achieved great success in supporting multimedia applications by understanding the alignments between images and text. While existing vision-language pretraining models primarily focus on understanding objects in images or entities in text, they often ignore the alignment at the level of events and their argument structures. In this work, we propose a contrastive learning framework to enforce vision-language pretraining models to comprehend events and associated argument (participant) roles. To achieve this, we take advantage of text information extraction technologies to obtain event structural knowledge, and utilize multiple prompt functions to contrast difficult negative descriptions by manipulating event structures. We also design an event graph alignment loss based on optimal transport to capture event argument structures. In addition, we collect a large event-rich dataset (106,875 images) for pretraining, which provides a more challenging image retrieval benchmark to assess the understanding of complicated lengthy sentences11The data and code are publicly available for research purpose in https://github.com/limanling/clip-event.. Experiments show that our zero-shot CLIP-Event outperforms the state-of-the-art supervised model in argument extraction on Multimedia Event Extraction, achieving more than 5% absolute F-score gain in event extraction, as well as significant improvements on a variety of downstream tasks under zero-shot settings.

Paraphrase identification has been formulated as a binary classification task to decide whether two sentences hold a paraphrase relationship.Existing paraphrase datasets only annotate a binary label for each sentence pair.However, after a systematical analysis of existing paraphrase datasets, we found that the degree of paraphrase cannot be well characterized by a single binary label.And the criteria of paraphrase are not even consistent within the same dataset.We hypothesize that such issues would limit the effectiveness of paraphrase models trained on these data.To this end, we propose a novel fine-grained paraphrase annotation schema that labels the minimum spans of tokens in a sentence that don't have the corresponding paraphrases in the other sentence.Under this setting, we frame paraphrasing as a sequence tagging task.We collect 30k sentence pairs in English with the new annotation schema, resulting in the ParaTag dataset.In addition to reporting baseline results on ParaTag using state-of-art language models, we show that ParaTag is especially useful for training an automatic scorer for language generation evaluation.Finally, we train a paraphrase generation model from ParaTag and achieve better data augmentation performance on the GLUE benchmark than other public paraphrasing datasets.1

The speech representations learned from large-scale unlabeled data have shown better generalizability than those from supervised learning and thus attract a lot of interest to be applied for various downstream tasks. In this paper, we explore the limits of speech representations learned by different self-supervised objectives and datasets for automatic speaker verification (ASV), especially with a well-recognized SOTA ASV model, ECAPA-TDNN [1], as a downstream model. The representations from all hidden layers of the pre-trained model are firstly averaged with learnable weights and then fed into the ECAPA-TDNN as input features. The experimental results on Voxceleb dataset show that the weighted average representation is significantly superior to FBank, a conventional handcrafted feature for ASV. Our best single system achieves 0.537%, 0.569%, and 1.180% equal error rate (EER) on the three official trials of VoxCeleb1, separately. Accordingly, the ensemble system with three pre-trained models can further improve the EER to 0.479%, 0.536% and 1.023%. Among the three evaluation trials, our best system outperforms the winner system [2] of the VoxCeleb Speaker Recognition Challenge 2021 (VoxSRC2021) on the VoxCeleb1-E trial.

The advances in attention-based encoder-decoder (AED) networks have brought great progress to end-to-end (E2E) automatic speech recognition (ASR). One way to further improve the performance of AED-based E2E ASR is to introduce an extra text encoder for leveraging extensive text data and thus capture more context-aware linguistic information. However, this approach brings a mismatch problem between the speech encoder and the text encoder due to the different units used for modeling. In this paper, we propose an embedding aligner and modality switch training to better align the speech and text latent spaces. The embedding aligner is a shared linear projection between text encoder and speech encoder trained by masked language modeling (MLM) loss and connectionist temporal classification (CTC), respectively. The modality switch training randomly swaps speech and text embeddings based on the forced alignment result to learn a joint representation space. Experimental results show that our proposed approach achieves a relative 14% to 19% word error rate (WER) reduction on Librispeech ASR task. We further verify its effectiveness on spoken language understanding (SLU), i.e., an absolute 2.5% to 2.8% F1 score improvement on SNIPS slot filling task.

Most of today's AI systems focus on using self-attention mechanisms and transformer architectures on large amounts of diverse data to achieve impressive performance gains. In this paper, we propose to augment the transformer architecture with an external attention mechanism to bring external knowledge and context to bear. By integrating external information into the prediction process, we hope to reduce the need for ever-larger models and increase the democratization of AI systems. We find that the proposed external attention mechanism can significantly improve the performance of existing AI systems, allowing practitioners to easily customize foundation AI models to many diverse downstream applications. In particular, we focus on the task of Commonsense Reasoning, demonstrating that the proposed external attention mechanism can augment existing transformer models and significantly improve the model's reasoning capabilities. The proposed system, Knowledgeable External Attention for commonsense Reasoning (KEAR), reaches human parity on the open CommonsenseQA research benchmark with an accuracy of 89.4% in comparison to the human accuracy of 88.9%.

People say, "A picture is worth a thousand words". Then how can we get the rich information out of the image? We argue that by using visual clues to bridge large pretrained vision foundation models and language models, we can do so without any extra cross-modal training. Thanks to the strong zero-shot capability of foundation models, we start by constructing a rich semantic representation of the image (e.g., image tags, object attributes / locations, captions) as a structured textual prompt, called visual clues, using a vision foundation model. Based on visual clues, we use large language model to produce a series of comprehensive descriptions for the visual content, which is then verified by the vision model again to select the candidate that aligns best with the image. We evaluate the quality of generated descriptions by quantitative and qualitative measurement. The results demonstrate the effectiveness of such a structured semantic representation.

In real application scenarios, it is often challenging to obtain a large amount of labeled data for speaker representation learning due to speaker privacy concerns. Self-supervised learning with no labels has become a more and more promising way to solve it. Compared with contrastive learning, self-distilled approaches use only positive samples in the loss function and thus are more attractive. In this paper, we present a comprehensive study on self-distilled self-supervised speaker representation learning, especially on critical data augmentation. Our proposed strategy of audio perturbation augmentation has pushed the performance of the speaker representation to a new limit. The experimental results show that our model can achieve a new SoTA on Voxceleb 1 speaker verification evaluation benchmark (i.e., equal error rate (EER) 2.505%, 2.473%, and 4.791 % for trial Vox1-O, Vox1-E and Vox1-H, respectively), discarding any speaker labels in the training phase.

Modern Automatic Speech Recognition (ASR) systems can achieve high performance in terms of recognition accuracy. However, a perfectly accurate transcript still can be challenging to read due to disfluency, filter words, and other errata common in spoken communication. Many downstream tasks and human readers rely on the output of the ASR system; therefore, errors introduced by the speaker and ASR system alike will be propagated to the next task in the pipeline. In this work, we propose an ASR post-processing model that aims to transform the incorrect and noisy ASR output into a readable text for humans and downstream tasks. We leverage the Metadata Extraction (MDE) corpus to construct a task-specific dataset for our study. Since the dataset is small, we propose a novel data augmentation method and use a two-stage training strategy to fine-tune the RoBERTa pre-trained model. On the constructed test set, our model outperforms a production two-step pipeline-based post-processing method by a large margin of 13.26 on readability-aware WER (RA-WER) and 17.53 on BLEU metrics. Human evaluation also demonstrates that our method can generate more human-readable transcripts than the baseline method.

End-to-end (E2E) spoken language understanding (SLU) can infer semantics directly from speech signal without cascading an automatic speech recognizer (ASR) with a natural language understanding (NLU) module. However, paired utterance recordings and corresponding semantics may not always be available or sufficient to train an E2E SLU model in a real production environment. In this paper, we propose to unify a well-optimized E2E ASR encoder (speech) and a pre-trained language model encoder (language) into a transformer decoder. The unified speech-language pre-trained model (SLP) is continually enhanced on limited labeled data from a target domain by using a conditional masked language model (MLM) objective, and thus can effectively generate a sequence of intent, slot type, and slot value for given input speech in the inference. The experimental results on two public corpora show that our approach to E2E SLU is superior to the conventional cascaded method. It also outperforms the present state-of-the-art approaches to E2E SLU with much less paired data.

In this paper, we propose a unified pre-training approach called UniSpeech to learn speech representations with both labeled and unlabeled data, in which supervised phonetic CTC learning and phonetically-aware contrastive self-supervised learning are conducted in a multi-task learning manner. The resultant representations can capture information more correlated with phonetic structures and improve the generalization across languages and domains. We evaluate the effectiveness of UniSpeech for cross-lingual representation learning on public CommonVoice corpus. The results show that UniSpeech outperforms self-supervised pretraining and supervised transfer learning for speech recognition by a maximum of 13.4% and 26.9% relative phone error rate reductions respectively (averaged over all testing languages). The transferability of UniSpeech is also verified on a domain-shift speech recognition task, i.e., a relative word error rate reduction of 6% against the previous approach.

Recently, universal neural machine translation (NMT) with shared encoder-decoder gained good performance on zero-shot translation. Unlike universal NMT, jointly trained language-specific encoders-decoders aim to achieve universal representation across non-shared modules, each of which is for a language or language family. The non-shared architecture has the advantage of mitigating internal language competition, especially when the shared vocabulary and model parameters are restricted in their size. However, the performance of using multiple encoders and decoders on zero-shot translation still lags behind universal NMT. In this work, we study zero-shot translation using language-specific encoders-decoders. We propose to generalize the non-shared architecture and universal NMT by differentiating the Transformer layers between language-specific and interlingua. By selectively sharing parameters and applying cross-attentions, we explore maximizing the representation universality and realizing the best alignment of language-agnostic information. We also introduce a denoising auto-encoding (DAE) objective to jointly train the model with the translation task in a multi-task manner. Experiments on two public multilingual parallel datasets show that our proposed model achieves competitive or better results than universal NMT and the strong pivot baseline. Moreover, we experiment incrementally adding new language to the trained model by only updating the new model parameters. With this little effort, the zero-shot translation between this newly added language and existing languages achieves a comparable result with the model trained jointly from scratch on all languages.

The training of spoken language understanding (SLU) models often faces the problem of data scarcity. In this paper, we put forward a data augmentation method using pretrained language models to boost the variability and accuracy of generated utterances. Furthermore, we investigate and propose solutions to two previously overlooked semi-supervised learning scenarios of data scarcity in SLU: i) Rich-in-Ontology: ontology information with numerous valid dialogue acts is given; ii) Rich-in-Utterance: a large number of unlabelled utterances are available. Empirical results show that our method can produce synthetic training data that boosts the performance of language understanding models in various scenarios.

Spoken language understanding (SLU) requires a model to analyze input acoustic signal to understand its linguistic content and make predictions.To boost the models' performance, various pre-training methods have been proposed to learn rich representations from large-scale unannotated speech and text.However, the inherent disparities between the two modalities necessitate a mutual analysis.In this paper, we propose a novel semisupervised learning framework, SPLAT, to jointly pre-train the speech and language modules.Besides conducting a self-supervised masked language modeling task on the two individual modules using unpaired speech and text, SPLAT aligns representations from the two modules in a shared latent space using a small amount of paired speech and text.Thus, during fine-tuning, the speech module alone can produce representations carrying both acoustic information and contextual semantic knowledge of an input acoustic signal.Experimental results verify the effectiveness of our approach on various SLU tasks.For example, SPLAT improves the previous stateof-the-art performance on the Spoken SQuAD dataset by more than 10%.

Chenguang Zhu, William Hinthorn, Ruochen Xu, Qingkai Zeng, Michael Zeng, Xuedong Huang, Meng Jiang. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

This paper introduces MEDIASUM 1 , a largescale media interview dataset consisting of 463.6K transcripts with abstractive summaries.To create this dataset, we collect interview transcripts from NPR and CNN and employ the overview and topic descriptions as summaries.Compared with existing public corpora for dialogue summarization, our dataset is an order of magnitude larger and contains complex multi-party conversations from multiple domains.We conduct statistical analysis to demonstrate the unique positional bias exhibited in the transcripts of televised and radioed interviews.We also show that MEDIASUM can be used in transfer learning to improve a model's performance on other dialogue summarization tasks.

A typical journalistic convention in news articles is to deliver the most salient information in the beginning, also known as the lead bias. While this phenomenon can be exploited in generating a summary, it has a detrimental effect on teaching a model to discriminate and extract important information in general. We propose that this lead bias can be leveraged in our favor in a simple and effective way to pre-train abstractive news summarization models on large-scale unlabeled news corpora: predicting the leading sentences using the rest of an article. We collect a massive news corpus and conduct data cleaning and filtering via statistical analysis. We then apply self-supervised pre-training on this dataset to existing generation models BART and T5 for domain adaptation. Via extensive experiments on six benchmark datasets, we show that this approach can dramatically improve the summarization quality and achieve state-of-the-art results for zero-shot news summarization without any fine-tuning. For example, in the DUC2003 dataset, the ROUGE-1 score of BART increases 13.7% after the lead-bias pre-training. We deploy the model in Microsoft News and provide public APIs as well as a demo website for multi-lingual news summarization.

In this work, we develop new self-learning techniques with an attention-based sequence-to-sequence (seq2seq) model for automatic speech recognition (ASR). For untranscribed speech data, the hypothesis from an ASR system must be used as a label. However, the imperfect ASR result makes unsupervised learning difficult to consistently improve recognition performance especially in the case that multiple powerful teacher models are unavailable. In contrast to conventional unsupervised learning approaches, we adopt the \emph{multi-task learning} (MTL) framework where the $n$-th best ASR hypothesis is used as the label of each task. The seq2seq network is updated through the MTL framework so as to find the common representation that can cover multiple hypotheses. By doing so, the effect of the \emph{hard-decision} errors can be alleviated. We first demonstrate the effectiveness of our self-learning methods through ASR experiments in an accent adaptation task between the US and British English speech. Our experiment results show that our method can reduce the WER on the British speech data from 14.55\% to 10.36\% compared to the baseline model trained with the US English data only. Moreover, we investigate the effect of our proposed methods in a federated learning scenario.

Chenguang Zhu, Michael Zeng, Xuedong Huang. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

We describe a system that generates speaker-annotated transcripts of meetings by using multiple asynchronous distant microphones. The system is composed of continuous audio stream alignment, blind beamforming, speech recognition, speaker diarization, and system combination. While the idea of improving the meeting transcription accuracy by leveraging multiple recordings has been investigated in certain specific technology areas such as beamforming, our objective is to assess the feasibility of a complete system with a set of mobile devices and conduct a detailed analysis. With seven input audio streams, our system achieves a word error rate (WER) of 22.3% and a speaker-attributed WER (SAWER) of 26.7%, and comes within 3% of the close-talking microphone WER on non-overlapping speech. The relative gains in SAWER over a single-device system are 14.8%, 20.3%, and 22.4% for three, five, and seven microphones, respectively.  The full system achieves a 13.6% diarization error rate, 10% of which are due to overlapped speech.

Dialogue state tracking is an important component in task-oriented dialogue systems to identify users' goals and requests as a dialogue proceeds.However, as most previous models are dependent on dialogue slots, the model complexity soars when the number of slots increases.In this paper, we put forward a slotindependent neural model (SIM) to track dialogue states while keeping the model complexity invariant to the number of dialogue slots.The model utilizes attention mechanisms between user utterance and system actions.SIM achieves state-of-the-art results on WoZ and DSTC2 tasks, with only 20% of the model size of previous models.