Yonghui Wu

Mingqiu Wang, Wei Han, Izhak Shafran et al. · deepmind

We present a joint Speech and Language Model (SLM), a multitask, multilingual, and dual-modal model that takes advantage of pretrained foundational speech and language models. SLM freezes the pretrained foundation models to maximally preserves their capabilities, and only trains a simple adapter with just 1\% (156M) of the foundation models' parameters. This adaptation not only leads SLM to achieve strong performance on conventional tasks such as speech recognition (ASR) and speech translation (AST), but also introduces the novel capability of zero-shot instruction-following for more diverse tasks: given a speech input and a text instruction, SLM is able to perform unseen generation tasks including contextual biasing ASR using real-time context, dialog generation, speech continuation, and question answering, etc. Our approach demonstrates that the representational gap between pretrained speech and language models might be narrower than one would expect, and can be bridged by a simple adaptation mechanism. As a result, SLM is not only efficient to train, but also inherits strong capabilities already acquired in foundation models of different modalities.

Zehao Yu, Xi Yang, Chong Dang et al.

Objective: We aim to develop an open-source natural language processing (NLP) package, SODA (i.e., SOcial DeterminAnts), with pre-trained transformer models to extract social determinants of health (SDoH) for cancer patients, examine the generalizability of SODA to a new disease domain (i.e., opioid use), and evaluate the extraction rate of SDoH using cancer populations. Methods: We identified SDoH categories and attributes and developed an SDoH corpus using clinical notes from a general cancer cohort. We compared four transformer-based NLP models to extract SDoH, examined the generalizability of NLP models to a cohort of patients prescribed with opioids, and explored customization strategies to improve performance. We applied the best NLP model to extract 19 categories of SDoH from the breast (n=7,971), lung (n=11,804), and colorectal cancer (n=6,240) cohorts. Results and Conclusion: We developed a corpus of 629 cancer patients notes with annotations of 13,193 SDoH concepts/attributes from 19 categories of SDoH. The Bidirectional Encoder Representations from Transformers (BERT) model achieved the best strict/lenient F1 scores of 0.9216 and 0.9441 for SDoH concept extraction, 0.9617 and 0.9626 for linking attributes to SDoH concepts. Fine-tuning the NLP models using new annotations from opioid use patients improved the strict/lenient F1 scores from 0.8172/0.8502 to 0.8312/0.8679. The extraction rates among 19 categories of SDoH varied greatly, where 10 SDoH could be extracted from >70% of cancer patients, but 9 SDoH had a low extraction rate (<70% of cancer patients). The SODA package with pre-trained transformer models is publicly available at https://github.com/uf-hobiinformatics-lab/SDoH_SODA.

Amin Dada, Aokun Chen, Cheng Peng et al.

Traditionally, large language models have been either trained on general web crawls or domain-specific data. However, recent successes of generative large language models, have shed light on the benefits of cross-domain datasets. To examine the significance of prioritizing data diversity over quality, we present a German dataset comprising texts from five domains, along with another dataset aimed at containing high-quality data. Through training a series of models ranging between 122M and 750M parameters on both datasets, we conduct a comprehensive benchmark on multiple downstream tasks. Our findings demonstrate that the models trained on the cross-domain dataset outperform those trained on quality data alone, leading to improvements up to $4.45\%$ over the previous state-of-the-art. The models are available at https://huggingface.co/ikim-uk-essen

Zihan Wang, Deli Chen, Damai Dai et al.

Parameter-efficient fine-tuning (PEFT) is crucial for customizing Large Language Models (LLMs) with constrained resources. Although there have been various PEFT methods for dense-architecture LLMs, PEFT for sparse-architecture LLMs is still underexplored. In this work, we study the PEFT method for LLMs with the Mixture-of-Experts (MoE) architecture and the contents of this work are mainly threefold: (1) We investigate the dispersion degree of the activated experts in customized tasks, and found that the routing distribution for a specific task tends to be highly concentrated, while the distribution of activated experts varies significantly across different tasks. (2) We propose Expert-Specialized Fine-Tuning, or ESFT, which tunes the experts most relevant to downstream tasks while freezing the other experts and modules; experimental results demonstrate that our method not only improves the tuning efficiency, but also matches or even surpasses the performance of full-parameter fine-tuning. (3) We further analyze the impact of the MoE architecture on expert-specialized fine-tuning. We find that MoE models with finer-grained experts are more advantageous in selecting the combination of experts that are most relevant to downstream tasks, thereby enhancing both the training efficiency and effectiveness. Our code is available at https://github.com/deepseek-ai/ESFT.

Yunshui Li, Yiyuan Ma, Shen Yan et al.

Model merging has emerged as a promising technique for enhancing large language models, though its application in large-scale pre-training remains relatively unexplored. In this paper, we present a comprehensive investigation of model merging techniques during the pre-training process. Through extensive experiments with both dense and Mixture-of-Experts (MoE) architectures ranging from millions to over 100 billion parameters, we demonstrate that merging checkpoints trained with constant learning rates not only achieves significant performance improvements but also enables accurate prediction of annealing behavior. These improvements lead to both more efficient model development and significantly lower training costs. Our detailed ablation studies on merging strategies and hyperparameters provide new insights into the underlying mechanisms while uncovering novel applications. Through comprehensive experimental analysis, we offer the open-source community practical pre-training guidelines for effective model merging.

Qianqian Xie, Qingyu Chen, Aokun Chen et al.

Recent advancements in large language models (LLMs) like ChatGPT and LLaMA show promise in medical applications, yet challenges remain in medical language comprehension. This study presents Me-LLaMA, a new medical LLM family based on open-source LLaMA models, optimized for medical text analysis and diagnosis by leveraging large-scale, domain-specific datasets. The Me-LLaMA family, including foundation models Me-LLaMA 13/70B and their chat-enhanced versions, was developed through continued pre-training and instruction tuning with 129B tokens and 214K samples from biomedical and clinical sources. Training the 70B models required over 100,000 A100 GPU hours. Me-LLaMA's performance was evaluated across six medical text analysis tasks using 12 benchmark datasets and complex clinical case diagnosis, with automatic and human evaluations. Results indicate Me-LLaMA outperforms LLaMA and other open-source medical LLMs in zero-shot and supervised settings. Task-specific tuning further boosts performance, surpassing ChatGPT on 7 of 8 datasets and GPT-4 on 5 of 8. For complex clinical cases, Me-LLaMA achieves performance comparable to ChatGPT and GPT-4. This work underscores the importance of domain-specific data in developing medical LLMs and addresses the high computational costs involved in training, highlighting a balance between pre-training and fine-tuning strategies. Me-LLaMA models are now accessible under user agreements, providing a valuable resource for advancing medical AI.

Yongye Su, Yuqing Wu

The ability of large language models to generate complex texts allows them to be widely integrated into many aspects of life, and their output can quickly fill all network resources. As the impact of LLMs grows, it becomes increasingly important to develop powerful detectors for the generated text. This detector is essential to prevent the potential misuse of these technologies and to protect areas such as social media from the negative effects of false content generated by LLMS. The main goal of LLM-generated text detection is to determine whether text is generated by an LLM, which is a basic binary classification task. In our work, we mainly use three different classification methods based on open source datasets: traditional machine learning techniques such as logistic regression, k-means clustering, Gaussian Naive Bayes, support vector machines, and methods based on converters such as BERT, and finally algorithms that use LLMs to detect LLM-generated text. We focus on model generalization, potential adversarial attacks, and accuracy of model evaluation. Finally, the possible research direction in the future is proposed, and the current experimental results are summarized.

Xi Yang, Zehao Yu, Yi Guo et al.

The newly emerged transformer technology has a tremendous impact on NLP research. In the general English domain, transformer-based models have achieved state-of-the-art performances on various NLP benchmarks. In the clinical domain, researchers also have investigated transformer models for clinical applications. The goal of this study is to systematically explore three widely used transformer-based models (i.e., BERT, RoBERTa, and XLNet) for clinical relation extraction and develop an open-source package with clinical pre-trained transformer-based models to facilitate information extraction in the clinical domain. We developed a series of clinical RE models based on three transformer architectures, namely BERT, RoBERTa, and XLNet. We evaluated these models using 2 publicly available datasets from 2018 MADE1.0 and 2018 n2c2 challenges. We compared two classification strategies (binary vs. multi-class classification) and investigated two approaches to generate candidate relations in different experimental settings. In this study, we compared three transformer-based (BERT, RoBERTa, and XLNet) models for relation extraction. We demonstrated that the RoBERTa-clinical RE model achieved the best performance on the 2018 MADE1.0 dataset with an F1-score of 0.8958. On the 2018 n2c2 dataset, the XLNet-clinical model achieved the best F1-score of 0.9610. Our results indicated that the binary classification strategy consistently outperformed the multi-class classification strategy for clinical relation extraction. Our methods and models are publicly available at https://github.com/uf-hobi-informatics-lab/ClinicalTransformerRelationExtraction. We believe this work will improve current practice on clinical relation extraction and other related NLP tasks in the biomedical domain.

Cheng Peng, Xi Yang, Aokun Chen et al.

Objective To solve major clinical natural language processing (NLP) tasks using a unified text-to-text learning architecture based on a generative large language model (LLM) via prompt tuning. Methods We formulated 7 key clinical NLP tasks as text-to-text learning and solved them using one unified generative clinical LLM, GatorTronGPT, developed using GPT-3 architecture and trained with up to 20 billion parameters. We adopted soft prompts (i.e., trainable vectors) with frozen LLM, where the LLM parameters were not updated (i.e., frozen) and only the vectors of soft prompts were updated, known as prompt tuning. We added additional soft prompts as a prefix to the input layer, which were optimized during the prompt tuning. We evaluated the proposed method using 7 clinical NLP tasks and compared them with previous task-specific solutions based on Transformer models. Results and Conclusion The proposed approach achieved state-of-the-art performance for 5 out of 7 major clinical NLP tasks using one unified generative LLM. Our approach outperformed previous task-specific transformer models by ~3% for concept extraction and 7% for relation extraction applied to social determinants of health, 3.4% for clinical concept normalization, 3.4~10% for clinical abbreviation disambiguation, and 5.5~9% for natural language inference. Our approach also outperformed a previously developed prompt-based machine reading comprehension (MRC) model, GatorTron-MRC, for clinical concept and relation extraction. The proposed approach can deliver the ``one model for all`` promise from training to deployment using a unified generative LLM.

Ricardo Loor-Torres, Yuqi Wu, Esteban Cabezas et al.

Background We aim to use Natural Language Processing (NLP) to automate the extraction and classification of thyroid cancer risk factors from pathology reports. Methods We analyzed 1,410 surgical pathology reports from adult papillary thyroid cancer patients at Mayo Clinic, Rochester, MN, from 2010 to 2019. Structured and non-structured reports were used to create a consensus-based ground truth dictionary and categorized them into modified recurrence risk levels. Non-structured reports were narrative, while structured reports followed standardized formats. We then developed ThyroPath, a rule-based NLP pipeline, to extract and classify thyroid cancer features into risk categories. Training involved 225 reports (150 structured, 75 unstructured), with testing on 170 reports (120 structured, 50 unstructured) for evaluation. The pipeline's performance was assessed using both strict and lenient criteria for accuracy, precision, recall, and F1-score. Results In extraction tasks, ThyroPath achieved overall strict F-1 scores of 93% for structured reports and 90 for unstructured reports, covering 18 thyroid cancer pathology features. In classification tasks, ThyroPath-extracted information demonstrated an overall accuracy of 93% in categorizing reports based on their corresponding guideline-based risk of recurrence: 76.9% for high-risk, 86.8% for intermediate risk, and 100% for both low and very low-risk cases. However, ThyroPath achieved 100% accuracy across all thyroid cancer risk categories with human-extracted pathology information. Conclusions ThyroPath shows promise in automating the extraction and risk recurrence classification of thyroid pathology reports at large scale. It offers a solution to laborious manual reviews and advancing virtual registries. However, it requires further validation before implementation.

Ziyi Chen, Mengyuan Zhang, Mustafa Mohammed Ahmed et al.

Cancer treatments are known to introduce cardiotoxicity, negatively impacting outcomes and survivorship. Identifying cancer patients at risk of heart failure (HF) is critical to improving cancer treatment outcomes and safety. This study examined machine learning (ML) models to identify cancer patients at risk of HF using electronic health records (EHRs), including traditional ML, Time-Aware long short-term memory (T-LSTM), and large language models (LLMs) using novel narrative features derived from the structured medical codes. We identified a cancer cohort of 12,806 patients from the University of Florida Health, diagnosed with lung, breast, and colorectal cancers, among which 1,602 individuals developed HF after cancer. The LLM, GatorTron-3.9B, achieved the best F1 scores, outperforming the traditional support vector machines by 39%, the T-LSTM deep learning model by 7%, and a widely used transformer model, BERT, by 5.6%. The analysis shows that the proposed narrative features remarkably increased feature density and improved performance.

Ahmad Idrissi-Yaghir, Amin Dada, Henning Schäfer et al.

Recent advances in natural language processing (NLP) can be largely attributed to the advent of pre-trained language models such as BERT and RoBERTa. While these models demonstrate remarkable performance on general datasets, they can struggle in specialized domains such as medicine, where unique domain-specific terminologies, domain-specific abbreviations, and varying document structures are common. This paper explores strategies for adapting these models to domain-specific requirements, primarily through continuous pre-training on domain-specific data. We pre-trained several German medical language models on 2.4B tokens derived from translated public English medical data and 3B tokens of German clinical data. The resulting models were evaluated on various German downstream tasks, including named entity recognition (NER), multi-label classification, and extractive question answering. Our results suggest that models augmented by clinical and translation-based pre-training typically outperform general domain models in medical contexts. We conclude that continuous pre-training has demonstrated the ability to match or even exceed the performance of clinical models trained from scratch. Furthermore, pre-training on clinical data or leveraging translated texts have proven to be reliable methods for domain adaptation in medical NLP tasks.

Yue wu

There are limitations of traditional methods and deep learning methods in terms of interpretability, generalization, and quantification of uncertainty in industrial fault diagnosis, and there are core problems of insufficient credibility in industrial fault diagnosis. The architecture performs preliminary analysis through a Bayesian network-based diagnostic engine and features an LLM-driven cognitive quorum module with multimodal input capabilities. The module conducts expert-level arbitration of initial diagnoses by analyzing structured features and diagnostic charts, prioritizing final decisions after conflicts are identified. To ensure the reliability of the system output, the architecture integrates a confidence calibration module based on temperature calibration and a risk assessment module, which objectively quantifies the reliability of the system using metrics such as expected calibration error (ECE). Experimental results on a dataset containing multiple fault types showed that the proposed framework improved diagnostic accuracy by more than 28 percentage points compared to the baseline model, while the calibrated ECE was reduced by more than 75%. Case studies have confirmed that HCAA effectively corrects misjudgments caused by complex feature patterns or knowledge gaps in traditional models, providing novel and practical engineering solutions for building high-trust, explainable AI diagnostic systems for industrial applications.

Cheng Peng, Xinyu Dong, Mengxian Lyu et al.

Natural language processing (NLP) is a key technology to extract important patient information from clinical narratives to support healthcare applications. The rapid development of large language models (LLMs) has revolutionized many NLP tasks in the clinical domain, yet their optimal use in patient information extraction tasks requires further exploration. This study examines LLMs' effectiveness in patient information extraction, focusing on LLM architectures, fine-tuning strategies, and multi-task instruction tuning techniques for developing robust and generalizable patient information extraction systems. This study aims to explore key concepts of using LLMs for clinical concept and relation extraction tasks, including: (1) encoder-only or decoder-only LLMs, (2) prompt-based parameter-efficient fine-tuning (PEFT) algorithms, and (3) multi-task instruction tuning on few-shot learning performance. We benchmarked a suite of LLMs, including encoder-based LLMs (BERT, GatorTron) and decoder-based LLMs (GatorTronGPT, Llama 3.1, GatorTronLlama), across five datasets. We compared traditional full-size fine-tuning and prompt-based PEFT. We explored a multi-task instruction tuning framework that combines both tasks across four datasets to evaluate the zero-shot and few-shot learning performance using the leave-one-dataset-out strategy.

Haofan Wu, Yin Huang, Yuqing Wu et al.

High-quality fundus images provide essential anatomical information for clinical screening and ophthalmic disease diagnosis. Yet, due to hardware limitations, operational variability, and patient compliance, fundus images often suffer from low resolution and signal-to-noise ratio. Recent years have witnessed promising progress in fundus image enhancement. However, existing works usually focus on restoring structural details or global characteristics of fundus images, lacking a unified image enhancement framework to recover comprehensive multi-scale information. Moreover, few methods pinpoint the target of image enhancement, e.g., lesions, which is crucial for medical image-based diagnosis. To address these challenges, we propose a multi-scale target-aware representation learning framework (MTRL-FIE) for efficient fundus image enhancement. Specifically, we propose a multi-scale feature encoder (MFE) that employs wavelet decomposition to embed both low-frequency structural information and high-frequency details. Next, we design a structure-preserving hierarchical decoder (SHD) to fuse multi-scale feature embeddings for real fundus image restoration. SHD integrates hierarchical fusion and group attention mechanisms to achieve adaptive feature fusion while retaining local structural smoothness. Meanwhile, a target-aware feature aggregation (TFA) module is used to enhance pathological regions and reduce artifacts. Experimental results on multiple fundus image datasets demonstrate the effectiveness and generalizability of MTRL-FIE for fundus image enhancement. Compared to state-of-the-art methods, MTRL-FIE achieves superior enhancement performance with a more lightweight architecture. Furthermore, our approach generalizes to other ophthalmic image processing tasks without supervised fine-tuning, highlighting its potential for clinical applications.

Ted Selker, Yunzi Wu

his paper explores what kinds of questions are best served by the way generative AI (GenAI) using Large Language Models(LLMs) that aggregate and package knowledge, and when traditional curated web-sourced search results serve users better. An experiment compared product searches using ChatGPT, Google search engine, or both helped us understand more about the compelling nature of generated responses. The experiment showed GenAI can speed up some explorations and decisions. We describe how search can deepen the testing of facts, logic, and context. We show where existing and emerging knowledge paradigms can help knowledge exploration in different ways. Experimenting with searches, our probes showed the value for curated web search provides for very specific, less popularly-known knowledge. GenAI excelled at bringing together knowledge for broad, relatively well-known topics. The value of curated and aggregated knowledge for different kinds of knowledge reflected in different user goals. We developed a taxonomy to distinguishing when users are best served by these two approaches.

Cheng Peng, Xi Yang, Aokun Chen et al.

There is enormous enthusiasm and concerns in using large language models (LLMs) in healthcare, yet current assumptions are all based on general-purpose LLMs such as ChatGPT. This study develops a clinical generative LLM, GatorTronGPT, using 277 billion words of mixed clinical and English text with a GPT-3 architecture of 20 billion parameters. GatorTronGPT improves biomedical natural language processing for medical research. Synthetic NLP models trained using GatorTronGPT generated text outperform NLP models trained using real-world clinical text. Physicians Turing test using 1 (worst) to 9 (best) scale shows that there is no significant difference in linguistic readability (p = 0.22; 6.57 of GatorTronGPT compared with 6.93 of human) and clinical relevance (p = 0.91; 7.0 of GatorTronGPT compared with 6.97 of human) and that physicians cannot differentiate them (p < 0.001). This study provides insights on the opportunities and challenges of LLMs for medical research and healthcare.

Jiahui Yu, Xin Li, Jing Yu Koh et al.

Pretraining language models with next-token prediction on massive text corpora has delivered phenomenal zero-shot, few-shot, transfer learning and multi-tasking capabilities on both generative and discriminative language tasks. Motivated by this success, we explore a Vector-quantized Image Modeling (VIM) approach that involves pretraining a Transformer to predict rasterized image tokens autoregressively. The discrete image tokens are encoded from a learned Vision-Transformer-based VQGAN (ViT-VQGAN). We first propose multiple improvements over vanilla VQGAN from architecture to codebook learning, yielding better efficiency and reconstruction fidelity. The improved ViT-VQGAN further improves vector-quantized image modeling tasks, including unconditional, class-conditioned image generation and unsupervised representation learning. When trained on ImageNet at \(256\times256\) resolution, we achieve Inception Score (IS) of 175.1 and Fr'echet Inception Distance (FID) of 4.17, a dramatic improvement over the vanilla VQGAN, which obtains 70.6 and 17.04 for IS and FID, respectively. Based on ViT-VQGAN and unsupervised pretraining, we further evaluate the pretrained Transformer by averaging intermediate features, similar to Image GPT (iGPT). This ImageNet-pretrained VIM-L significantly beats iGPT-L on linear-probe accuracy from 60.3% to 73.2% for a similar model size. VIM-L also outperforms iGPT-XL which is trained with extra web image data and larger model size.

Yu Zhang, Daniel S. Park, Wei Han et al.

We summarize the results of a host of efforts using giant automatic speech recognition (ASR) models pre-trained using large, diverse unlabeled datasets containing approximately a million hours of audio. We find that the combination of pre-training, self-training and scaling up model size greatly increases data efficiency, even for extremely large tasks with tens of thousands of hours of labeled data. In particular, on an ASR task with 34k hours of labeled data, by fine-tuning an 8 billion parameter pre-trained Conformer model we can match state-of-the-art (SoTA) performance with only 3% of the training data and significantly improve SoTA with the full training set. We also report on the universal benefits gained from using big pre-trained and self-trained models for a large set of downstream tasks that cover a wide range of speech domains and span multiple orders of magnitudes of dataset sizes, including obtaining SoTA performance on many public benchmarks. In addition, we utilize the learned representation of pre-trained networks to achieve SoTA results on non-ASR tasks.

Jeffrey Zhao, Mahdis Mahdieh, Ye Zhang et al.

Sequence-to-sequence models have been applied to a wide variety of NLP tasks, but how to properly use them for dialogue state tracking has not been systematically investigated. In this paper, we study this problem from the perspectives of pre-training objectives as well as the formats of context representations. We demonstrate that the choice of pre-training objective makes a significant difference to the state tracking quality. In particular, we find that masked span prediction is more effective than auto-regressive language modeling. We also explore using Pegasus, a span prediction-based pre-training objective for text summarization, for the state tracking model. We found that pre-training for the seemingly distant summarization task works surprisingly well for dialogue state tracking. In addition, we found that while recurrent state context representation works also reasonably well, the model may have a hard time recovering from earlier mistakes. We conducted experiments on the MultiWOZ 2.1-2.4, WOZ 2.0, and DSTC2 datasets with consistent observations.

Zehao Yu, Xi Yang, Chong Dang et al.

Social and behavioral determinants of health (SBDoH) have important roles in shaping people's health. In clinical research studies, especially comparative effectiveness studies, failure to adjust for SBDoH factors will potentially cause confounding issues and misclassification errors in either statistical analyses and machine learning-based models. However, there are limited studies to examine SBDoH factors in clinical outcomes due to the lack of structured SBDoH information in current electronic health record (EHR) systems, while much of the SBDoH information is documented in clinical narratives. Natural language processing (NLP) is thus the key technology to extract such information from unstructured clinical text. However, there is not a mature clinical NLP system focusing on SBDoH. In this study, we examined two state-of-the-art transformer-based NLP models, including BERT and RoBERTa, to extract SBDoH concepts from clinical narratives, applied the best performing model to extract SBDoH concepts on a lung cancer screening patient cohort, and examined the difference of SBDoH information between NLP extracted results and structured EHRs (SBDoH information captured in standard vocabularies such as the International Classification of Diseases codes). The experimental results show that the BERT-based NLP model achieved the best strict/lenient F1-score of 0.8791 and 0.8999, respectively. The comparison between NLP extracted SBDoH information and structured EHRs in the lung cancer patient cohort of 864 patients with 161,933 various types of clinical notes showed that much more detailed information about smoking, education, and employment were only captured in clinical narratives and that it is necessary to use both clinical narratives and structured EHRs to construct a more complete picture of patients' SBDoH factors.

Yu-An Chung, Yu Zhang, Wei Han et al.

Motivated by the success of masked language modeling~(MLM) in pre-training natural language processing models, we propose w2v-BERT that explores MLM for self-supervised speech representation learning. w2v-BERT is a framework that combines contrastive learning and MLM, where the former trains the model to discretize input continuous speech signals into a finite set of discriminative speech tokens, and the latter trains the model to learn contextualized speech representations via solving a masked prediction task consuming the discretized tokens. In contrast to existing MLM-based speech pre-training frameworks such as HuBERT, which relies on an iterative re-clustering and re-training process, or vq-wav2vec, which concatenates two separately trained modules, w2v-BERT can be optimized in an end-to-end fashion by solving the two self-supervised tasks~(the contrastive task and MLM) simultaneously. Our experiments show that w2v-BERT achieves competitive results compared to current state-of-the-art pre-trained models on the LibriSpeech benchmarks when using the Libri-Light~60k corpus as the unsupervised data. In particular, when compared to published models such as conformer-based wav2vec~2.0 and HuBERT, our model shows~5\% to~10\% relative WER reduction on the test-clean and test-other subsets. When applied to the Google's Voice Search traffic dataset, w2v-BERT outperforms our internal conformer-based wav2vec~2.0 by more than~30\% relatively.

Yuanzhong Xu, HyoukJoong Lee, Dehao Chen et al.

We present GSPMD, an automatic, compiler-based parallelization system for common machine learning computations. It allows users to write programs in the same way as for a single device, then give hints through a few annotations on how to distribute tensors, based on which GSPMD will parallelize the computation. Its representation of partitioning is simple yet general, allowing it to express different or mixed paradigms of parallelism on a wide variety of models. GSPMD infers the partitioning for every operator based on limited user annotations, making it convenient to scale existing single-device programs. It solves several technical challenges for production usage, allowing GSPMD to achieve 50% to 62% compute utilization on up to 2048 Cloud TPUv3 cores for models with up to one trillion parameters.

Ye Jia, Heiga Zen, Jonathan Shen et al.

This paper introduces PnG BERT, a new encoder model for neural TTS. This model is augmented from the original BERT model, by taking both phoneme and grapheme representations of text as input, as well as the word-level alignment between them. It can be pre-trained on a large text corpus in a self-supervised manner, and fine-tuned in a TTS task. Experimental results show that a neural TTS model using a pre-trained PnG BERT as its encoder yields more natural prosody and more accurate pronunciation than a baseline model using only phoneme input with no pre-training. Subjective side-by-side preference evaluations show that raters have no statistically significant preference between the speech synthesized using a PnG BERT and ground truth recordings from professional speakers.

Isaac Elias, Heiga Zen, Jonathan Shen et al.

This paper introduces Parallel Tacotron 2, a non-autoregressive neural text-to-speech model with a fully differentiable duration model which does not require supervised duration signals. The duration model is based on a novel attention mechanism and an iterative reconstruction loss based on Soft Dynamic Time Warping, this model can learn token-frame alignments as well as token durations automatically. Experimental results show that Parallel Tacotron 2 outperforms baselines in subjective naturalness in several diverse multi speaker evaluations. Its duration control capability is also demonstrated.

Ye Zhang, Yuan Cao, Mahdis Mahdieh et al.

Dialogue state tracking (DST) is a pivotal component in task-oriented dialogue systems. While it is relatively easy for a DST model to capture belief states in short conversations, the task of DST becomes more challenging as the length of a dialogue increases due to the injection of more distracting contexts. In this paper, we aim to improve the overall performance of DST with a special focus on handling longer dialogues. We tackle this problem from three perspectives: 1) A model designed to enable hierarchical slot status prediction; 2) Balanced training procedure for generic and task-specific language understanding; 3) Data perturbation which enhances the model's ability in handling longer conversations. We conduct experiments on the MultiWOZ benchmark, and demonstrate the effectiveness of each component via a set of ablation tests, especially on longer conversations.

Walker Ravina, Ethan Sterling, Olexiy Oryeshko et al.

The goal of model distillation is to faithfully transfer teacher model knowledge to a model which is faster, more generalizable, more interpretable, or possesses other desirable characteristics. Human-readability is an important and desirable standard for machine-learned model interpretability. Readable models are transparent and can be reviewed, manipulated, and deployed like traditional source code. As a result, such models can be improved outside the context of machine learning and manually edited if desired. Given that directly training such models is difficult, we propose to train interpretable models using conventional methods, and then distill them into concise, human-readable code. The proposed distillation methodology approximates a model's univariate numerical functions with piecewise-linear curves in a localized manner. The resulting curve model representations are accurate, concise, human-readable, and well-regularized by construction. We describe a piecewise-linear curve-fitting algorithm that produces high-quality results efficiently and reliably across a broad range of use cases. We demonstrate the effectiveness of the overall distillation technique and our curve-fitting algorithm using four datasets across the tasks of classification, regression, and ranking.

Bo Li, Anmol Gulati, Jiahui Yu et al.

End-to-end (E2E) models have shown to outperform state-of-the-art conventional models for streaming speech recognition [1] across many dimensions, including quality (as measured by word error rate (WER)) and endpointer latency [2]. However, the model still tends to delay the predictions towards the end and thus has much higher partial latency compared to a conventional ASR model. To address this issue, we look at encouraging the E2E model to emit words early, through an algorithm called FastEmit [3]. Naturally, improving on latency results in a quality degradation. To address this, we explore replacing the LSTM layers in the encoder of our E2E model with Conformer layers [4], which has shown good improvements for ASR. Secondly, we also explore running a 2nd-pass beam search to improve quality. In order to ensure the 2nd-pass completes quickly, we explore non-causal Conformer layers that feed into the same 1st-pass RNN-T decoder, an algorithm called Cascaded Encoders [5]. Overall, we find that the Conformer RNN-T with Cascaded Encoders offers a better quality and latency tradeoff for streaming ASR.

Isaac Elias, Heiga Zen, Jonathan Shen et al.

Although neural end-to-end text-to-speech models can synthesize highly natural speech, there is still room for improvements to its efficiency and naturalness. This paper proposes a non-autoregressive neural text-to-speech model augmented with a variational autoencoder-based residual encoder. This model, called \emph{Parallel Tacotron}, is highly parallelizable during both training and inference, allowing efficient synthesis on modern parallel hardware. The use of the variational autoencoder relaxes the one-to-many mapping nature of the text-to-speech problem and improves naturalness. To further improve the naturalness, we use lightweight convolutions, which can efficiently capture local contexts, and introduce an iterative spectrogram loss inspired by iterative refinement. Experimental results show that Parallel Tacotron matches a strong autoregressive baseline in subjective evaluations with significantly decreased inference time.

Jiahui Yu, Chung-Cheng Chiu, Bo Li et al.

Streaming automatic speech recognition (ASR) aims to emit each hypothesized word as quickly and accurately as possible. However, emitting fast without degrading quality, as measured by word error rate (WER), is highly challenging. Existing approaches including Early and Late Penalties and Constrained Alignments penalize emission delay by manipulating per-token or per-frame probability prediction in sequence transducer models. While being successful in reducing delay, these approaches suffer from significant accuracy regression and also require additional word alignment information from an existing model. In this work, we propose a sequence-level emission regularization method, named FastEmit, that applies latency regularization directly on per-sequence probability in training transducer models, and does not require any alignment. We demonstrate that FastEmit is more suitable to the sequence-level optimization of transducer models for streaming ASR by applying it on various end-to-end streaming ASR networks including RNN-Transducer, Transformer-Transducer, ConvNet-Transducer and Conformer-Transducer. We achieve 150-300 ms latency reduction with significantly better accuracy over previous techniques on a Voice Search test set. FastEmit also improves streaming ASR accuracy from 4.4%/8.9% to 3.1%/7.5% WER, meanwhile reduces 90th percentile latency from 210 ms to only 30 ms on LibriSpeech.

Yu Zhang, James Qin, Daniel S. Park et al.

We employ a combination of recent developments in semi-supervised learning for automatic speech recognition to obtain state-of-the-art results on LibriSpeech utilizing the unlabeled audio of the Libri-Light dataset. More precisely, we carry out noisy student training with SpecAugment using giant Conformer models pre-trained using wav2vec 2.0 pre-training. By doing so, we are able to achieve word-error-rates (WERs) 1.4%/2.6% on the LibriSpeech test/test-other sets against the current state-of-the-art WERs 1.7%/3.3%.

Jiahui Yu, Wei Han, Anmol Gulati et al.

Streaming automatic speech recognition (ASR) aims to emit each hypothesized word as quickly and accurately as possible, while full-context ASR waits for the completion of a full speech utterance before emitting completed hypotheses. In this work, we propose a unified framework, Dual-mode ASR, to train a single end-to-end ASR model with shared weights for both streaming and full-context speech recognition. We show that the latency and accuracy of streaming ASR significantly benefit from weight sharing and joint training of full-context ASR, especially with inplace knowledge distillation during the training. The Dual-mode ASR framework can be applied to recent state-of-the-art convolution-based and transformer-based ASR networks. We present extensive experiments with two state-of-the-art ASR networks, ContextNet and Conformer, on two datasets, a widely used public dataset LibriSpeech and a large-scale dataset MultiDomain. Experiments and ablation studies demonstrate that Dual-mode ASR not only simplifies the workflow of training and deploying streaming and full-context ASR models, but also significantly improves both emission latency and recognition accuracy of streaming ASR. With Dual-mode ASR, we achieve new state-of-the-art streaming ASR results on both LibriSpeech and MultiDomain in terms of accuracy and latency.

Jonathan Shen, Ye Jia, Mike Chrzanowski et al.

This paper presents Non-Attentive Tacotron based on the Tacotron 2 text-to-speech model, replacing the attention mechanism with an explicit duration predictor. This improves robustness significantly as measured by unaligned duration ratio and word deletion rate, two metrics introduced in this paper for large-scale robustness evaluation using a pre-trained speech recognition model. With the use of Gaussian upsampling, Non-Attentive Tacotron achieves a 5-scale mean opinion score for naturalness of 4.41, slightly outperforming Tacotron 2. The duration predictor enables both utterance-wide and per-phoneme control of duration at inference time. When accurate target durations are scarce or unavailable in the training data, we propose a method using a fine-grained variational auto-encoder to train the duration predictor in a semi-supervised or unsupervised manner, with results almost as good as supervised training.

Daniel S. Park, Yu Zhang, Ye Jia et al.

Recently, a semi-supervised learning method known as "noisy student training" has been shown to improve image classification performance of deep networks significantly. Noisy student training is an iterative self-training method that leverages augmentation to improve network performance. In this work, we adapt and improve noisy student training for automatic speech recognition, employing (adaptive) SpecAugment as the augmentation method. We find effective methods to filter, balance and augment the data generated in between self-training iterations. By doing so, we are able to obtain word error rates (WERs) 4.2%/8.6% on the clean/noisy LibriSpeech test sets by only using the clean 100h subset of LibriSpeech as the supervised set and the rest (860h) as the unlabeled set. Furthermore, we are able to achieve WERs 1.7%/3.4% on the clean/noisy LibriSpeech test sets by using the unlab-60k subset of LibriLight as the unlabeled set for LibriSpeech 960h. We are thus able to improve upon the previous state-of-the-art clean/noisy test WERs achieved on LibriSpeech 100h (4.74%/12.20%) and LibriSpeech (1.9%/4.1%).

Anmol Gulati, James Qin, Chung-Cheng Chiu et al.

Recently Transformer and Convolution neural network (CNN) based models have shown promising results in Automatic Speech Recognition (ASR), outperforming Recurrent neural networks (RNNs). Transformer models are good at capturing content-based global interactions, while CNNs exploit local features effectively. In this work, we achieve the best of both worlds by studying how to combine convolution neural networks and transformers to model both local and global dependencies of an audio sequence in a parameter-efficient way. To this regard, we propose the convolution-augmented transformer for speech recognition, named Conformer. Conformer significantly outperforms the previous Transformer and CNN based models achieving state-of-the-art accuracies. On the widely used LibriSpeech benchmark, our model achieves WER of 2.1%/4.3% without using a language model and 1.9%/3.9% with an external language model on test/testother. We also observe competitive performance of 2.7%/6.3% with a small model of only 10M parameters.

Aditya Siddhant, Ankur Bapna, Yuan Cao et al.

Over the last few years two promising research directions in low-resource neural machine translation (NMT) have emerged. The first focuses on utilizing high-resource languages to improve the quality of low-resource languages via multilingual NMT. The second direction employs monolingual data with self-supervision to pre-train translation models, followed by fine-tuning on small amounts of supervised data. In this work, we join these two lines of research and demonstrate the efficacy of monolingual data with self-supervision in multilingual NMT. We offer three major results: (i) Using monolingual data significantly boosts the translation quality of low-resource languages in multilingual models. (ii) Self-supervision improves zero-shot translation quality in multilingual models. (iii) Leveraging monolingual data with self-supervision provides a viable path towards adding new languages to multilingual models, getting up to 33 BLEU on ro-en translation without any parallel data or back-translation.

Chung-Cheng Chiu, Arun Narayanan, Wei Han et al.

In recent years, all-neural end-to-end approaches have obtained state-of-the-art results on several challenging automatic speech recognition (ASR) tasks. However, most existing works focus on building ASR models where train and test data are drawn from the same domain. This results in poor generalization characteristics on mismatched-domains: e.g., end-to-end models trained on short segments perform poorly when evaluated on longer utterances. In this work, we analyze the generalization properties of streaming and non-streaming recurrent neural network transducer (RNN-T) based end-to-end models in order to identify model components that negatively affect generalization performance. We propose two solutions: combining multiple regularization techniques during training, and using dynamic overlapping inference. On a long-form YouTube test set, when the nonstreaming RNN-T model is trained with shorter segments of data, the proposed combination improves word error rate (WER) from 22.3% to 14.8%; when the streaming RNN-T model trained on short Search queries, the proposed techniques improve WER on the YouTube set from 67.0% to 25.3%. Finally, when trained on Librispeech, we find that dynamic overlapping inference improves WER on YouTube from 99.8% to 33.0%.

Wei Han, Zhengdong Zhang, Yu Zhang et al.

Convolutional neural networks (CNN) have shown promising results for end-to-end speech recognition, albeit still behind other state-of-the-art methods in performance. In this paper, we study how to bridge this gap and go beyond with a novel CNN-RNN-transducer architecture, which we call ContextNet. ContextNet features a fully convolutional encoder that incorporates global context information into convolution layers by adding squeeze-and-excitation modules. In addition, we propose a simple scaling method that scales the widths of ContextNet that achieves good trade-off between computation and accuracy. We demonstrate that on the widely used LibriSpeech benchmark, ContextNet achieves a word error rate (WER) of 2.1%/4.6% without external language model (LM), 1.9%/4.1% with LM and 2.9%/7.0% with only 10M parameters on the clean/noisy LibriSpeech test sets. This compares to the previous best published system of 2.0%/4.6% with LM and 3.9%/11.3% with 20M parameters. The superiority of the proposed ContextNet model is also verified on a much larger internal dataset.

Honglei Zhuang, Xuanhui Wang, Michael Bendersky et al.

Interpretability of learning-to-rank models is a crucial yet relatively under-examined research area. Recent progress on interpretable ranking models largely focuses on generating post-hoc explanations for existing black-box ranking models, whereas the alternative option of building an intrinsically interpretable ranking model with transparent and self-explainable structure remains unexplored. Developing fully-understandable ranking models is necessary in some scenarios (e.g., due to legal or policy constraints) where post-hoc methods cannot provide sufficiently accurate explanations. In this paper, we lay the groundwork for intrinsically interpretable learning-to-rank by introducing generalized additive models (GAMs) into ranking tasks. Generalized additive models (GAMs) are intrinsically interpretable machine learning models and have been extensively studied on regression and classification tasks. We study how to extend GAMs into ranking models which can handle both item-level and list-level features and propose a novel formulation of ranking GAMs. To instantiate ranking GAMs, we employ neural networks instead of traditional splines or regression trees. We also show that our neural ranking GAMs can be distilled into a set of simple and compact piece-wise linear functions that are much more efficient to evaluate with little accuracy loss. We conduct experiments on three data sets and show that our proposed neural ranking GAMs can achieve significantly better performance than other traditional GAM baselines while maintaining similar interpretability.

Tara N. Sainath, Yanzhang He, Bo Li et al.

Thus far, end-to-end (E2E) models have not been shown to outperform state-of-the-art conventional models with respect to both quality, i.e., word error rate (WER), and latency, i.e., the time the hypothesis is finalized after the user stops speaking. In this paper, we develop a first-pass Recurrent Neural Network Transducer (RNN-T) model and a second-pass Listen, Attend, Spell (LAS) rescorer that surpasses a conventional model in both quality and latency. On the quality side, we incorporate a large number of utterances across varied domains to increase acoustic diversity and the vocabulary seen by the model. We also train with accented English speech to make the model more robust to different pronunciations. In addition, given the increased amount of training data, we explore a varied learning rate schedule. On the latency front, we explore using the end-of-sentence decision emitted by the RNN-T model to close the microphone, and also introduce various optimizations to improve the speed of LAS rescoring. Overall, we find that RNN-T+LAS offers a better WER and latency tradeoff compared to a conventional model. For example, for the same latency, RNN-T+LAS obtains a 8% relative improvement in WER, while being more than 400-times smaller in model size.

Guangzhi Sun, Yu Zhang, Ron J. Weiss et al.

This paper proposes a hierarchical, fine-grained and interpretable latent variable model for prosody based on the Tacotron 2 text-to-speech model. It achieves multi-resolution modeling of prosody by conditioning finer level representations on coarser level ones. Additionally, it imposes hierarchical conditioning across all latent dimensions using a conditional variational auto-encoder (VAE) with an auto-regressive structure. Evaluation of reconstruction performance illustrates that the new structure does not degrade the model while allowing better interpretability. Interpretations of prosody attributes are provided together with the comparison between word-level and phone-level prosody representations. Moreover, both qualitative and quantitative evaluations are used to demonstrate the improvement in the disentanglement of the latent dimensions.

Guangzhi Sun, Yu Zhang, Ron J. Weiss et al.

Recent neural text-to-speech (TTS) models with fine-grained latent features enable precise control of the prosody of synthesized speech. Such models typically incorporate a fine-grained variational autoencoder (VAE) structure, extracting latent features at each input token (e.g., phonemes). However, generating samples with the standard VAE prior often results in unnatural and discontinuous speech, with dramatic prosodic variation between tokens. This paper proposes a sequential prior in a discrete latent space which can generate more naturally sounding samples. This is accomplished by discretizing the latent features using vector quantization (VQ), and separately training an autoregressive (AR) prior model over the result. We evaluate the approach using listening tests, objective metrics of automatic speech recognition (ASR) performance, and measurements of prosody attributes. Experimental results show that the proposed model significantly improves the naturalness in random sample generation. Furthermore, initial experiments demonstrate that randomly sampling from the proposed model can be used as data augmentation to improve the ASR performance.

Daniel S. Park, Yu Zhang, Chung-Cheng Chiu et al.

Recently, SpecAugment, an augmentation scheme for automatic speech recognition that acts directly on the spectrogram of input utterances, has shown to be highly effective in enhancing the performance of end-to-end networks on public datasets. In this paper, we demonstrate its effectiveness on tasks with large scale datasets by investigating its application to the Google Multidomain Dataset (Narayanan et al., 2018). We achieve improvement across all test domains by mixing raw training data augmented with SpecAugment and noise-perturbed training data when training the acoustic model. We also introduce a modification of SpecAugment that adapts the time mask size and/or multiplicity depending on the length of the utterance, which can potentially benefit large scale tasks. By using adaptive masking, we are able to further improve the performance of the Listen, Attend and Spell model on LibriSpeech to 2.2% WER on test-clean and 5.2% WER on test-other.

Chung-Cheng Chiu, Wei Han, Yu Zhang et al.

End-to-end automatic speech recognition (ASR) models, including both attention-based models and the recurrent neural network transducer (RNN-T), have shown superior performance compared to conventional systems. However, previous studies have focused primarily on short utterances that typically last for just a few seconds or, at most, a few tens of seconds. Whether such architectures are practical on long utterances that last from minutes to hours remains an open question. In this paper, we both investigate and improve the performance of end-to-end models on long-form transcription. We first present an empirical comparison of different end-to-end models on a real world long-form task and demonstrate that the RNN-T model is much more robust than attention-based systems in this regime. We next explore two improvements to attention-based systems that significantly improve its performance: restricting the attention to be monotonic, and applying a novel decoding algorithm that breaks long utterances into shorter overlapping segments. Combining these two improvements, we show that attention-based end-to-end models can be very competitive to RNN-T on long-form speech recognition.

Xi Yang, Yan Gong, Nida Waheed et al.

Cardiotoxicity related to cancer therapies has become a serious issue, diminishing cancer treatment outcomes and quality of life. Early detection of cancer patients at risk for cardiotoxicity before cardiotoxic treatments and providing preventive measures are potential solutions to improve cancer patients's quality of life. This study focuses on predicting the development of heart failure in cancer patients after cancer diagnoses using historical electronic health record (EHR) data. We examined four machine learning algorithms using 143,199 cancer patients from the University of Florida Health (UF Health) Integrated Data Repository (IDR). We identified a total number of 1,958 qualified cases and matched them to 15,488 controls by gender, age, race, and major cancer type. Two feature encoding strategies were compared to encode variables as machine learning features. The gradient boosting (GB) based model achieved the best AUC score of 0.9077 (with a sensitivity of 0.8520 and a specificity of 0.8138), outperforming other machine learning methods. We also looked into the subgroup of cancer patients with exposure to chemotherapy drugs and observed a lower specificity score (0.7089). The experimental results show that machine learning methods are able to capture clinical factors that are known to be associated with heart failure and that it is feasible to use machine learning methods to identify cancer patients at risk for cancer therapy-related heart failure.

Andrew Rosenberg, Yu Zhang, Bhuvana Ramabhadran et al.

Recent success of the Tacotron speech synthesis architecture and its variants in producing natural sounding multi-speaker synthesized speech has raised the exciting possibility of replacing expensive, manually transcribed, domain-specific, human speech that is used to train speech recognizers. The multi-speaker speech synthesis architecture can learn latent embedding spaces of prosody, speaker and style variations derived from input acoustic representations thereby allowing for manipulation of the synthesized speech. In this paper, we evaluate the feasibility of enhancing speech recognition performance using speech synthesis using two corpora from different domains. We explore algorithms to provide the necessary acoustic and lexical diversity needed for robust speech recognition. Finally, we demonstrate the feasibility of this approach as a data augmentation strategy for domain-transfer. We find that improvements to speech recognition performance is achievable by augmenting training data with synthesized material. However, there remains a substantial gap in performance between recognizers trained on human speech those trained on synthesized speech.

Anjuli Kannan, Arindrima Datta, Tara N. Sainath et al.

Multilingual end-to-end (E2E) models have shown great promise in expansion of automatic speech recognition (ASR) coverage of the world's languages. They have shown improvement over monolingual systems, and have simplified training and serving by eliminating language-specific acoustic, pronunciation, and language models. This work presents an E2E multilingual system which is equipped to operate in low-latency interactive applications, as well as handle a key challenge of real world data: the imbalance in training data across languages. Using nine Indic languages, we compare a variety of techniques, and find that a combination of conditioning on a language vector and training language-specific adapter layers produces the best model. The resulting E2E multilingual model achieves a lower word error rate (WER) than both monolingual E2E models (eight of nine languages) and monolingual conventional systems (all nine languages).

Tara N. Sainath, Ruoming Pang, David Rybach et al.

The requirements for many applications of state-of-the-art speech recognition systems include not only low word error rate (WER) but also low latency. Specifically, for many use-cases, the system must be able to decode utterances in a streaming fashion and faster than real-time. Recently, a streaming recurrent neural network transducer (RNN-T) end-to-end (E2E) model has shown to be a good candidate for on-device speech recognition, with improved WER and latency metrics compared to conventional on-device models [1]. However, this model still lags behind a large state-of-the-art conventional model in quality [2]. On the other hand, a non-streaming E2E Listen, Attend and Spell (LAS) model has shown comparable quality to large conventional models [3]. This work aims to bring the quality of an E2E streaming model closer to that of a conventional system by incorporating a LAS network as a second-pass component, while still abiding by latency constraints. Our proposed two-pass model achieves a 17%-22% relative reduction in WER compared to RNN-T alone and increases latency by a small fraction over RNN-T.

Naveen Arivazhagan, Ankur Bapna, Orhan Firat et al.

We introduce our efforts towards building a universal neural machine translation (NMT) system capable of translating between any language pair. We set a milestone towards this goal by building a single massively multilingual NMT model handling 103 languages trained on over 25 billion examples. Our system demonstrates effective transfer learning ability, significantly improving translation quality of low-resource languages, while keeping high-resource language translation quality on-par with competitive bilingual baselines. We provide in-depth analysis of various aspects of model building that are crucial to achieving quality and practicality in universal NMT. While we prototype a high-quality universal translation system, our extensive empirical analysis exposes issues that need to be further addressed, and we suggest directions for future research.

Yu Zhang, Ron J. Weiss, Heiga Zen et al.

We present a multispeaker, multilingual text-to-speech (TTS) synthesis model based on Tacotron that is able to produce high quality speech in multiple languages. Moreover, the model is able to transfer voices across languages, e.g. synthesize fluent Spanish speech using an English speaker's voice, without training on any bilingual or parallel examples. Such transfer works across distantly related languages, e.g. English and Mandarin. Critical to achieving this result are: 1. using a phonemic input representation to encourage sharing of model capacity across languages, and 2. incorporating an adversarial loss term to encourage the model to disentangle its representation of speaker identity (which is perfectly correlated with language in the training data) from the speech content. Further scaling up the model by training on multiple speakers of each language, and incorporating an autoencoding input to help stabilize attention during training, results in a model which can be used to consistently synthesize intelligible speech for training speakers in all languages seen during training, and in native or foreign accents.