检索结果-内蒙古大学图书馆

IEEE Spoken Language Technology Workshop (SLT)

作者： Lei, Yi Yang, Shan Xie, Lei Northwestern Polytech Univ Sch Comp Sci Audio Speech & Language Proc Grp ASLP NPU Xian Peoples R China

ISBN: (纸本)9781728170664

This paper proposes a unified model to conduct emotion transfer, control and prediction for sequence-to-sequence based fine-grained emotional speech synthesis. Conventional emotional speech synthesis often needs manual labels or reference audio to determine the emotional expressions of synthesized speech. Such coarse labels cannot control the details of speech emotion, often resulting in an averaged emotion expression delivery, and it is also hard to choose suitable reference audio during inference. To conduct fine-grained emotion expression generation, we introduce phoneme-level emotion strength representations through a learned ranking function to describe the local emotion details, and the sentence-level emotion category is adopted to render the global emotions of synthesized speech. With the global render and local descriptors of emotions, we can obtain fine-grained emotion expressions from reference audio via its emotion descriptors (for transfer) or directly from phoneme-level manual labels (for control). As for the emotional speech synthesis with arbitrary text inputs, the proposed model can also predict phoneme-level emotion expressions from texts, which does not require any reference audio or manual label.

关键词： text-to-speech expressive speech synthesis emotion strength sequence-to-sequence

来源：评论

学校读者我要写书评

暂无评论

PATNET : A PHONEME-LEVEL AUTOREGRESSIVE TRANSFORMER NETWORK FOR SPEECH SYNTHESIS

PATNET : A PHONEME-LEVEL AUTOREGRESSIVE TRANSFORMER NETWORK ...

引用

IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

作者： Wang, Shiming Ling, Zhenhua Fu, Ruibo Yi, Jiangyan Tao, Jianhua Univ Sci & Technol China NELSLIP Hefei Peoples R China Chinese Acad Sci Inst Automat NLPR Beijing Peoples R China

ISBN: (纸本)9781728176055

Aiming at efficiently predicting acoustic features with high naturalness and robustness, this paper proposes PATNet, a neural acoustic model for speech synthesis using phoneme-level autoregression. PATNet accepts phoneme sequences as input and is built based on Transformer structure. PATNet adopts a duration model instead of attention mechanism for sequence alignment. The decoder of PATNet predicts multi-frame spectra within one phoneme in parallel given the predicted spectra of previous phonemes. Such phoneme-level autoregression enables PATNet to achieve higher inference efficiency than the models with frame-level autoregression, such as Transformer-TTS, and improves the robustness of acoustic feature prediction by utilizing phoneme boundaries explicitly. Experimental results show that the speech synthesized by PATNet obtained lower character error rate (CER) than Tacotron, Transfomer-TTS and FastSpeech when evaluated by a speech recognition engine. Besides, PATNet achieved 10 times faster inference speed than TransformerTTS and significantly better naturalness than FastSpeech.

关键词： speech synthesis sequence-to-sequence Transformer phoneme-level autoregression

来源：评论

学校读者我要写书评

暂无评论

RECONSTRUCTING DUAL LEARNING FOR NEURAL VOICE CONVERSION USING RELATIVELY FEW SAMPLES

RECONSTRUCTING DUAL LEARNING FOR NEURAL VOICE CONVERSION USI...

引用

IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)

作者： Sun, Aolan Wang, Jianzong Cheng, Ning Tantrawenith, Methawee Wu, Zhiyong Meng, Helen Xiao, Edward Xiao, Jing Ping An Technol Shenzhen Co Shenzhen Peoples R China Tsinghua Univ Tsinghua Shenzhen Int Grad Sch Shenzhen Peoples R China Chinese Univ Hong Kong Hong Kong Peoples R China Aquinas Int Acad La Palma CA USA

ISBN: (纸本)9781665437394

This paper introduces a dual learning system for neural voice conversion (DualVC) using relatively few samples based on the symmetry of the speech conversion task. The system contains a pair of sequence-to-sequence neural networks that have the same structure but are trained in opposite directions. The objective function of the dual model training is the sum of paired conversion loss and reconstruction loss during the dual training circle. The models in the two directions are trained alternately to guide each other by the corresponding reconstruction loss. Furthermore, curriculum learning techniques are used to load models in existing fields into the current task to accelerate the rapid iteration and convergence of the model. The experiment on the voice conversion task with the proposed DualVC and curriculum learning strategy obtained a comparable naturalness and similarity with only a 30% dataset than the BaseVC model trained on the full dataset.

关键词： speech synthesis voice conversion dual learning sequence-to-sequence curriculum learning

来源：评论

学校读者我要写书评

暂无评论

Phonetically Induced Subwords for End-to-End Speech Recognition 22

Phonetically Induced Subwords for End-to-End Speech Recognit...

引用

Interspeech Conference

作者： Papadourakis, Vasileios Mueller, Markus Liu, Jing Mouchtaris, Athanasios Omologo, Maurizio Amazon Alexa Machine Learning Seattle WA 98109 USA

ISBN: (纸本)9781713836902

End-to-end automatic speech recognition systems map a sequence of acoustic features to text. In modern systems, text is encoded to grapheme subwords which are generated by methods designed for text processing tasks and therefore don't model or take advantage of the statistics of the acoustic features. Here, we present a novel method for generating grapheme subwords that are derived from phoneme sequences, therefore capturing phonetical statistics. The phonetically induced subwords can be used for training and inference in any system that benefits from subwords, regardless of architecture and without the need of a pronunciation lexicon. We compare our method to other commonly used methods, which are based on text statistics or on text-phoneme correspondence and present experiments on CTC and RNN-T architectures, evaluating subword sets of different sizes. We find that our phonetically induced subwords can improve performance of RNN-T models with relative improvements of up to 15.21% compared to other subword methods.

关键词： word piece RNN-T CTC sequence-to-sequence phoneme

来源：评论

学校读者我要写书评

暂无评论

FINE-TUNING OF PRE-TRAINED END-TO-END SPEECH RECOGNITION WITH GENERATIVE ADVERSARIAL NETWORKS

FINE-TUNING OF PRE-TRAINED END-TO-END SPEECH RECOGNITION WIT...

引用

IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

作者： Haidar, Md Akmal Rezagholizadeh, Mehdi Montreal Res Ctr Huawei Noahs Ark Lab Montreal PQ Canada

ISBN: (纸本)9781728176055

Adversarial training of end-to-end (E2E) ASR systems using generative adversarial networks (GAN) has recently been explored for low-resource ASR corpora. GANs help to learn the true data representation through a two-player min-max game. However, training an E2E ASR model using a large ASR corpus with a GAN framework has never been explored, because it might take excessively long time due to high-variance gradient updates and face convergence issues. In this paper, we introduce a novel framework for fine-tuning a pre-trained ASR model using the GAN objective where the ASR model acts as a generator and a discriminator tries to distinguish the ASR output from the real data. Since the ASR model is pre-trained, we hypothesize that the ASR model output (soft distribution vectors) helps to get higher scores from the discriminator and makes the task of the discriminator harder within our GAN framework, which in turn improves the performance of the ASR model in the fine-tuning stage. Here, the pre-trained ASR model is fine-tuned adversarially against the discriminator using an additional adversarial loss. Experiments on full LibriSpeech dataset show that our proposed approach outperforms baselines and conventional GAN-based adversarial models.

关键词： automatic speech recognition sequence-to-sequence transformer generative adversarial networks adversarial training

来源：评论

学校读者我要写书评

暂无评论

Investigating Recurrent Neural Networks for Diacritizing Arabic Text and Correcting Soft Spelling Mistakes

Investigating Recurrent Neural Networks for Diacritizing Ara...

引用

IEEE Jordan International Joint Conference on Electrical Engineering and Information Technology (JEEIT)

作者： Almajdoubah, Ahmad N. Abandah, Gheith A. Suyyagh, Ashraf E. Univ Jordan Comp Engn Dept Amman Jordan

ISBN: (纸本)9781665442930

Modern Arabic language processing research aims to find efficient solutions to important problems such as adding diacritics to undiacritized text and correcting spelling mistakes. Some previous machine learning solutions solve these problems as one-to-one sequence transcription problems, which imposes restrictions on the type of errors that can be corrected. This work investigates transformer and encoder-decoder recurrent neural network models taking these two problems as case studies. These models solve sequence to sequence transcription problems without restrictions on the input and output lengths. We investigated several alternatives and evaluated them on two benchmark datasets. We recommend an encoder-decoder model that provides excellent accuracy. The encoder part has an embedding layer and a long short-term memory (LSTM) layer. The decoder part has an embedding layer, an LSTM layer, Luong attention between the two parts, and a dense output layer with softmax activation.

关键词： encoder-decoder recurrent neural network attention sequence-to-sequence Arabic diacritization spelling mistakes

来源：评论

学校读者我要写书评

暂无评论

Multilingual fine-tuning for Grammatical Error Correction

引用

EXPERT SYSTEMS WITH APPLICATIONS 2022年第0期200卷 1页

作者： Pajak, Krzysztof Pajak, Dominik LangMedia Sp Z OO Mariana Rapackiego 5 PL-53021 Wroclaw Poland Wroclaw Univ Sci & Technol Wybrzeze Wyspianskiego 27 PL-50370 Wroclaw Poland

Finding a single model capable of comprehending multiple languages is an area of active research in Natural Language Processing (NLP). Recently developed models such as mBART, mT5 or xProphetNet can solve problems connected with, for instance, machine translation and summarization for many languages. However, good multilingual solutions to the problem of Grammatical Error Correction (GEC) are still missing - this paper aims at filling this gap. We first review current annotated GEC datasets and then apply existing pre-trained multilingual models to correct grammatical errors in multiple languages. In our experiments, we compare how different pre-training approaches impact the final GEC quality. Our result is a single model that can correct seven different languages and is the best (in terms of F-score) currently reported multilingual GEC model. Additionally, our multilingual model achieves better results than the SOTA monolingual model for Romanian.

关键词： Multilingual denoising autoencoder transformer sequence-to-sequence Grammatical error correction

来源：评论

学校读者我要写书评

暂无评论

End-to-End Detection and Recognition of Arithmetic Expressions 4th

End-to-End Detection and Recognition of Arithmetic Expressio...

引用

4th Chinese Conference on Pattern Recognition and Computer Vision (PRCV)

作者： Wan, Jiangpeng Zhao, Mengbiao Yin, Fei Zhang, Xu-Yao Huang, LinLin Beijing Jiaotong Univ Beijing 100044 Peoples R China Chinese Acad Sci Natl Lab Pattern Recognit Inst Automat Beijing 100190 Peoples R China Univ Chinese Acad Sci Sch Artificial Intelligence Beijing 100049 Peoples R China

ISBN: (纸本)9783030880040;9783030880033

The detection and recognition of handwritten arithmetic expressions (AEs) play an important role in document retrieval [ 21] and analysis. They are very difficult because of the structural complexity and the variability of appearance. In this paper, we propose a novel framework to detect and recognize AEs in an End-to-End manner. Firstly, an AE detector based on EfficientNet-B1 [17] is designed to locate all AE instances efficiently. Upon AE location, the RoI Rotate module [11] is adopted to transform visual features for AE proposals. The transformed features are then fed into an attention mechanism based recognizer for AE recognition. The whole network for detection and recognition is trained End-to-End on document images annotated AE locations and transcripts. Since the datasets in this field are rare, we also construct a dataset named HAED, which contains 1069 images (855 for training, and 214 for testing). Extensive experiments on two datasets (HAED and TFD-ICDAR 2019) show that the proposed method has achieved competitive performance on both datasets.

关键词： Arithmetic expression spotting End-to-End training sequence-to-sequence

来源：评论

学校读者我要写书评

暂无评论

Deep reinforcement and transfer learning for abstractive text summarization: A review

引用

COMPUTER SPEECH AND LANGUAGE 2022年 71卷 101276-101276页

作者： Alomari, Ayham Idris, Norisma Sabri, Aznul Qalid Md Alsmadi, Izzat Univ Malaya Fac Comp Sci & Informat Technol Dept Artificial Intelligence Kuala Lumpur 50603 Malaysia Texas A&M Dept Comp & Cyber Secur San Antonio TX USA

Automatic Text Summarization (ATS) is an important area in Natural Language Processing (NLP) with the goal of shortening a long text into a more compact version by conveying the most important points in a readable form. ATS applications continue to evolve and utilize effective approaches that are being evaluated and implemented by researchers. State-of-the-Art (SotA) technologies that demonstrate cutting-edge performance and accuracy in abstractive ATS are deep neural sequence-to-sequence models, Reinforcement Learning (RL) approaches, and Transfer Learning (TL) approaches, including Pre-Trained Language Models (PTLMs). The graph-based Transformer architecture and PTLMs have influenced tremendous advances in NLP applications. Additionally, the incorporation of recent mechanisms, such as the knowledge-enhanced mechanism, significantly enhanced the results. This study provides a comprehensive review of recent research advances in the area of abstractive text summarization for works spanning the past six years. Past and present problems are described, as well as their proposed solutions. In addition, abstractive ATS datasets and evaluation measurements are also highlighted. The paper concludes by comparing the best models and discussing future research directions.

关键词： Abstractive summarization sequence-to-sequence Reinforcement learning Pre-trained models

来源：评论

学校读者我要写书评

暂无评论

On the role of the architecture for spring discharge prediction with deep learning approaches

引用

HYDROLOGICAL PROCESSES 2022年第10期36卷 e14737-e14737页

作者： Zhou, Renjie Zhang, Yanyan Sam Houston State Univ Dept Environm & Geosci Huntsville TX 77340 USA Texas A&M Univ Dept Elect & Comp Engn College Stn TX 77840 USA

Understanding karst spring flow is important to accommodate the increasing water demand caused by the population growth and manage the freshwater water resource effectively. However, due to the spatial and temporal heterogeneity and complex hydrological processes in karst systems, predicting karst spring discharge remains challenging. In this study, three deep learning-based models, including long short-term memory (LSTM), gated recurrent unit (GRU) and simple recurrent neural network (RNN), are framed with an encoder-decoder architecture to provide multiple-step-ahead spring discharge prediction. The encoder-decoder architecture includes an encoder that reads and encodes the input sequence into a vector and decoder that deciphers the vector and outputs the predicted sequence. Three hybrid models called LSTM-ED, GRU-ED and simple RNN-ED are compared with single-step models and multiple-step models without the encoder-decoder architecture to investigate the role of the encoder-decoder architecture on multi-step-ahead prediction. The sensitivity of the selection of input time and lead time steps on the karst spring discharge prediction is evaluated. The predicted results are compared with the observed spring discharge. It implies that: (1) LSTM-ED, GRU-ED and RNN-ED models obtain similar results on predicting karst spring discharge multiple time steps ahead;(2) three hybrid multiple-step models outperform the single-step models in making consistent and accurate spring discharge predictions;(3) the multiple-step models framed with an encoder-decoder architecture obtain better spring discharge prediction results than the single-step models and multiple-step models without the encoder-decoder structure;(4) the LSTM-ED, GRU-ED and simple RNN-ED models are sensitive to the selection of lead time and insensitive to the selection of input time step. A short lead time typically yields a more accurate spring discharge prediction.

关键词： encoder-decoder GRU karst LSTM RNN sequence-to-sequence spring discharge

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：