End-to-end multilingual ASR has become more appealing because of several reasons such as simplifying the training and deployment process and positive performance transfer from high-resource to low-resource languages. However, scaling up the number of languages, total hours, and number of unique tokens is not a trivial task. This paper explores large-scale multilingual ASR models on 70 languages. We inspect two architectures: (1) Shared embedding and output and (2) Multiple embedding and output model. In the shared model experiments, we show the importance of tokenization strategy across different languages. Later, we use our optimal tokenization strategy to train multiple embedding and output model to further improve our result. Our multilingual ASR achieves 13.9%-15.6% average WER relative improvement compared to monolingual models. We show that our multilingual ASR generalizes well on an unseen dataset and domain, achieving 9.5% and 7.5% WER on Multilingual Librispeech (MLS) with zero-shot and finetuning, respectively.
translated by 谷歌翻译
神经网络修剪可以有效地用于压缩自动语音识别(ASR)模型。但是,在多语言ASR中,执行语言不足的修剪可能会导致某些语言的严重性能降解,因为语言 - 敏捷的修剪口罩可能不符合所有语言,并丢弃了重要的语言特定参数。在这项工作中,我们提出了ASR路径,这是一种稀疏的多语言ASR模型,该模型激活了特定语言的子网络(“路径”),从而明确地学习了每种语言的参数。通过重叠的子网络,共享参数还可以通过联合多语言培训来实现较低资源语言的知识传输。我们提出了一种新型算法来学习ASR途径,并通过流式RNN-T模型评估了4种语言的建议方法。我们提出的ASR途径的表现都优于密集模型(平均-5.0%)和语言不足的修剪模型(平均-21.4%),并且与单语稀疏模型相比,低资源语言的性能更好。
translated by 谷歌翻译
越来越有兴趣将流和全文自动语音识别(ASR)网络统一到单个端到端ASR模型中,以简化两种用例的模型培训和部署。在现实世界中的ASR应用程序中,流媒体ASR模型通常在更多的存储和计算约束(例如,在嵌入式设备上)进行操作,而不是任何服务器端的全文模型。由Omni-Sparsity Supernet训练的最新进展激发,该训练在一个单个模型中共同优化了多个子网,该工作旨在共同学习紧凑的稀疏稀疏式磁性流媒体流动ASR模型,以及一个大型密度服务器非流动模型,在一个超级网。接下来,我们提出,在两种WAV2VEC 2.0自制学习和监督的ASR微调上进行超网训练不仅可以基本上改善先前工作中所示的大型非流式模型,还可以改善紧凑的稀疏流流媒体流模型。
translated by 谷歌翻译
我们提出了一种基于审议的新型方法来端到端(E2E)口语理解(SLU),其中流媒体自动语音识别(ASR)模型会产生第一频繁的假设和第二通通的自然语言(NLU)(NLU) )组件通过对ASR的文本和音频嵌入来生成语义解析。通过将E2E SLU制定为广义解码器,我们的系统能够支持复杂的组成语义结构。此外,ASR和NLU之间的参数共享使该系统特别适合资源受限的(内部设备)环境;我们提出的方法始终在TOPV2数据集的口头版本(Stop)的口语版本上始终优于强大管道NLU基线的0.60%至0.65%。我们证明了文本和音频功能的融合,再加上系统重写第一通道假设的能力,使我们的方法对ASR错误更加强大。最后,我们表明我们的方法可以显着减少从自然语音到合成语音训练时的降解,但是要使文本到语音(TTS)成为可行的解决方案,以扩大E2E SLU。
translated by 谷歌翻译
450万小时的英语演讲从10个不同的10个不同来源,跨越高达10亿参数的不同来源,我们探索了自动语音识别的规模前沿。我们提出了数据选择技术,以有效地缩放培训数据,以找到大规模数据集中最有价值的样本。为了有效地进行模型尺寸,我们利用各种优化,例如稀疏传感器丢失和模型分片。通过培训1-10B参数通用英语ASR模型,我们将语音识别性能的限制推动在许多域中。此外,我们的模型学习强大的语音表示,在新域名和言语方面具有零和少量功能,超出了多个内部和公共基准的先前结果。对于由于脑损伤而具有障碍的扬声器,我们最好的零射击和少量射频分别在Aphasiabank测试集中实现了22%和60%,同时在公共社交媒体视频中实现了最佳性能。此外,相同的通用模型在SPGISPeech Financial-Domain数据集上达到了500倍的域内数据等效性能。
translated by 谷歌翻译
从可穿戴设备到功能强大的智能设备,现代自动语音识别(ASR)型号在各种具有不同计算预算的边缘设备上运行。为了浏览模型准确性与模型大小的帕累托前线,研究人员陷入了通过为每个单独的边缘设备进行训练和微调模型来优化模型精度的困境,同时保持训练GPU小时可拖动。在本文中,我们提出了Omni-Sparsity DNN,其中可以修剪单个神经网络以生成针对各种模型大小的优化模型。我们为Omni-Sparsity DNN制定了培训策略,使其可以在Word-Error-rate(WER)vs模型大小的帕累托(Pareto)沿线找到模​​型,同时使培训GPU小时不超过训练一个单数模型的模型。我们使用流e2e ASR模型演示了Omni-Sparsity DNN。与单独修剪的稀疏型号相比,我们的结果在LibrisPeech上具有相似或更高准确性的培训时间和资源节省了大量节省:在测试中差2%-6.6%。
translated by 谷歌翻译
测量自动语音识别(ASR)系统质量对于创建用户满意的语音驱动应用程序至关重要。传统上,单词错误率(WER)用于评估ASR系统质量;但是,它有时与用户对转录质量的看法/判断息息相关。这是因为wer平等地称重每个单词,并且不考虑对用户感知产生更高影响的语义正确性。在这项工作中,我们提出评估ASR输出的质量,可以通过使用参考的语义向量与从预训练的语言模型中提取的假设之间的距离来测量语义正确性。我们对71K和36K用户注释的ASR输出质量的实验结果表明,与WER相比,Semdist与用户感知的相关性更高。我们还表明,与WER相比,Semdist与下游自然语言理解(NLU)任务具有更高的相关性。
translated by 谷歌翻译
Although machine learning (ML) models of AI achieve high performances in medicine, they are not free of errors. Empowering clinicians to identify incorrect model recommendations is crucial for engendering trust in medical AI. Explainable AI (XAI) aims to address this requirement by clarifying AI reasoning to support the end users. Several studies on biomedical imaging achieved promising results recently. Nevertheless, solutions for models using tabular data are not sufficient to meet the requirements of clinicians yet. This paper proposes a methodology to support clinicians in identifying failures of ML models trained with tabular data. We built our methodology on three main pillars: decomposing the feature set by leveraging clinical context latent space, assessing the clinical association of global explanations, and Latent Space Similarity (LSS) based local explanations. We demonstrated our methodology on ML-based recognition of preterm infant morbidities caused by infection. The risk of mortality, lifelong disability, and antibiotic resistance due to model failures was an open research question in this domain. We achieved to identify misclassification cases of two models with our approach. By contextualizing local explanations, our solution provides clinicians with actionable insights to support their autonomy for informed final decisions.
translated by 谷歌翻译
Objective: Social Determinants of Health (SDOH) influence personal health outcomes and health systems interactions. Health systems capture SDOH information through structured data and unstructured clinical notes; however, clinical notes often contain a more comprehensive representation of several key SDOH. The objective of this work is to assess the SDOH information gain achievable by extracting structured semantic representations of SDOH from the clinical narrative and combining these extracted representations with available structured data. Materials and Methods: We developed a natural language processing (NLP) information extraction model for SDOH that utilizes a deep learning entity and relation extraction architecture. In an electronic health record (EHR) case study, we applied the SDOH extractor to a large existing clinical data set with over 200,000 patients and 400,000 notes and compared the extracted information with available structured data. Results: The SDOH extractor achieved 0.86 F1 on a withheld test set. In the EHR case study, we found 19\% of current tobacco users, 10\% of drug users, and 32\% of homeless patients only include documentation of these risk factors in the clinical narrative. Conclusions: Patients who are at-risk for negative health outcomes due to SDOH may be better served if health systems are able to identify SDOH risk factors and associated social needs. Structured semantic representations of text-encoded SDOH information can augment existing structured, and this more comprehensive SDOH representation can assist health systems in identifying and addressing social needs.
translated by 谷歌翻译
随着计算机视觉应用程序的最新增长,尚未探索它们的公平和公正性问题。有大量证据表明,训练数据中存在的偏差反映在模型中,甚至放大。图像数据集的许多以前的方法偏见,包括基于增强数据集的模型,在计算上实现的计算昂贵。在这项研究中,我们提出了一个快速有效的模型,以通过重建并最大程度地减少预期变量之间的统计依赖性来消除图像数据集。我们的体系结构包括重建图像的U-NET,并结合了预先训练的分类器,该分类器会惩罚目标属性和受保护属性之间的统计依赖性。我们在Celeba数据集上评估了我们提出的模型,将结果与最先进的偏见方法进行比较,并证明该模型实现了有希望的公平性 - 精确性组合。
translated by 谷歌翻译