DOCSLIB.ORG

End-To-End Simultaneous Speech to Text Translation

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
End-To-End Simultaneous Speech to Text Translation SimulSpeech: End-to-End Simultaneous Speech to Text Translation Yi Ren ∗ Jinglin Liu ∗ Xu Tan ∗ Zhejiang University Zhejiang University Microsoft Research [email protected] [email protected] [email protected] Chen Zhang ∗ Tao Qin Zhou Zhaoy Zhejiang University Microsoft Research Zhejiang University [email protected] [email protected] [email protected] Tie-Yan Liu Microsoft Research [email protected] Abstract 1 Introduction Simultaneous speech to text translation (Fugen¨ In this work, we develop SimulSpeech, an end- et al., 2007; Oda et al., 2014; Dalvi et al., 2018), to-end simultaneous speech to text translation system which translates speech in source lan- which translates source-language speech into target- guage to text in target language concurrently. language text concurrently, is of great importance SimulSpeech consists of a speech encoder, a to the real-time understanding of spoken lectures or speech segmenter and a text decoder, where conversations and now widely used in many scenar- 1) the segmenter builds upon the encoder and ios including live video streaming and international leverages a connectionist temporal classifica- conferences. However, it is widely considered as tion (CTC) loss to split the input streaming one of the challenging tasks in machine transla- speech in real time, 2) the encoder-decoder at- tention adopts a wait-k strategy for simulta- tion domain because simultaneous speech to text neous translation. SimulSpeech is more chal- translation has to understand the speech and trade lenging than previous cascaded systems (with off translation accuracy and delay. Conventional simultaneous automatic speech recognition approaches to simultaneous speech to text transla- (ASR) and simultaneous neural machine trans- tion (Fugen¨ et al., 2007; Oda et al., 2014; Dalvi lation (NMT)). We introduce two novel knowl- et al., 2018) divide the translation process into two edge distillation methods to ensure the perfor- stages: simultaneous automatic speech recognition mance: 1) Attention-level knowledge distilla- (ASR) (Rao et al., 2017) and simultaneous neu- tion transfers the knowledge from the multipli- cation of the attention matrices of simultane- ral machine translation (NMT) (Gu et al., 2016), ous NMT and ASR models to help the training which cannot be optimized jointly and result in in- of the attention mechanism in SimulSpeech; 2) ferior accuracy, and also incurs more translation Data-level knowledge distillation transfers the delay due to two stages. knowledge from the full-sentence NMT model In this paper, we move a step further to translate and also reduces the complexity of data distri- the source speech to target text simultaneously, and bution to help on the optimization of Simul- Speech. Experiments on MuST-C English- develop SimulSpeech, an end-to-end simultaneous Spanish and English-German spoken language speech to text translation system. The SimulSpeech translation datasets show that SimulSpeech model consists of 1) a speech encoder where each achieves reasonable BLEU scores and lower speech frame can only see its previous frames to delay compared to full-sentence end-to-end simulate streaming speech inputs; 2) a text decoder speech to text translation (without simultane- where the encoder-decoder attention follows the ous translation), and better performance than wait-k strategy (Ma et al., 2018) to decide when to the two-stage cascaded simultaneous transla- tion model in terms of BLEU scores and trans- listen and write on the source speech and target text lation delay. respectively (see Figure1); 3) a speech segmenter that builds upon the encoder and leverages a CTC ∗ Equal contribution. loss to detect the word boundary, which is used y Corresponding author to decide when to stop listening according to the 3787 Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3787–3796 July 5 - 10, 2020. c 2020 Association for Computational Linguistics Source Audio → 2 Preliminaries In this section, we briefly review some basic knowl- edge for simultaneous speech to text translation, How including speech to text translation, simultaneous translation based on wait-k strategy, and CTC loss Target text text Target is Listen for segmentation. Write the → Speech to Text Translation Given a set of bilin- weather gual speech-text sentence pairs D = f(x; y) 2 (X × Y)g, an speech to text machine translation today model learns the parameter θ by minimizing the P negative log-likelihood − (x;y)2D log P (yjx; θ). Figure 1: The wait-k strategy for simultaneous speech P (yjx; θ) is calculated based on the chain rule QTy to text translation. The model will wait for the first t=1 P (ytjy<t; x; θ), where y<t represents the k source speech segments and then start to translate a text tokens preceding position t, and Ty is the target word. After that, once receiving a new source length of text sentence y. An encoder-attention- segment, the decoder generates a new target word until decoder framework is usually adopted to model there is no more source word, and then the translation the conditional probability P (yjx; θ), where the degrades to the full-sentence translation. The example shows the case with k = 2. encoder maps the input audio to a set of hidden representations h and the decoder generates each target token yt using the previously generated to- wait-k strategy. kens y<t as well as the speech representations h. Considering the difficulty of this task, we elab- Previous works (Berard´ et al., 2016; Weiss et al., orately design two techniques to boost the perfor- 2017; Liu et al., 2019) on speech to text translation mance of SimulSpeech: 1) attention-level knowl- focus on the full-sentence translation where the full edge distillation that transfers the knowledge from source speech can be seen when predicting each the multiplication of the attention matrices of si- target token. multaneous NMT and ASR model to SimulSpeech Simultaneous Translation Based on Wait-k Si- to help the training of its attention mechanism; 2) multaneous translation aims to translate sentences data-level knowledge distillation that transfers the before they are finished according to certain strate- knowledge from a full-sentence NMT model to gies. We use wait-k strategy (Ma et al., 2018) SimulSpeech and also reduces the complexity of in this work: given a set of speech and text data distribution (Zhou et al., 2019) to help on the pairs D = f(x; y) 2 (X × Y)g, the model optimization of SimulSpeech model. with the wait-k strategy learns the parameter θ Compared with the cascaded pipeline that trains by minimizing the negative log-likelihood loss P simultaneous ASR and NMT models separately, − (x;y)2D log P (yjx; k; θ), where k corresponds SimulSpeech can alleviate the error propagation to the wait-k strategy. P (yjx; k; θ) is calculated problem and optimize all model parameters jointly based on the chain rule towards the end goal, as well as reduce the de- Ty lay of simultaneous translation. Experiments on Y P (yjx; k; θ) = P (ytjy<t; x ; θ); (1) 1 <t+k MuST-C English-Spanish and English-German t=1 spoken language translation datasets demonstrate that SimulSpeech: 1) achieves reasonable BLEU where y<t represents the tokens preceding position scores and lower delay compared to full-sentence t and Ty is the length of target sentence y, x<t+k end-to-end speech to text translation (without si- represents the speech segments preceding position multaneous translation), and 2) obtains better per- t + k. The wait-k strategy ensures that the model formance than the two-stage cascaded simultane- can see t + k − 1 source segments when generat- ous translation model in terms of BLEU scores and ing the target token yt, while can see the whole translation delay. sentence if there is no more source segments. CTC for Alignment and Segmentation The 1https://ict.fbk.eu/must-c/ connectionist temporal classification (CTC) 3788 Output text Softmax Linear Word boundary Speech Add & Norm Segmenter Feed Forward Data Distillation Add & Norm Source Text Encoder Decoder Add & Norm (Source Text) (Target Text) Wait-K Decoder Auxiliary Tasks Feed Forward Attention Attention Distillation Source Text NMT Attention ASR Attention N× Add & Norm Add & Norm ×N Auxiliary NMT Auxiliary ASR x Predicted Encoder Decoder Source Masked Self- Masked Self- (source text, masked) (source text, wait-k) Text attention Attention CTC loss Speech Segmenter + Decoder Predicted Encoder (target text, wait-k) (source speech, masked) Target S2T Attention Text Pre-Net Word Embedding Source Speech Target text (a) SimulSpeech model structure. (b) The training pipeline for SimulSpeech model. Figure 2: (a) The model structure of SimulSpeech. (b) The training pipeline for SimulSpeech model. The Simul- Speech model is shown in purple box, and the auxiliary training techniques are in other boxes. loss (Graves et al., 2006) is widely used for features, which consists of multiple convo- alignment and segmentation, which maps the lutional layers with the same hidden size as frame-level classification outputs of a speech Transformer. sequence to a text sequence (with a different length from the speech sequence). For a text sequence • To enable simultaneous translation, we design y, CTC introduces a set of intermediate represen- different attention mechanisms for the encoder tation paths φ(y) called CTC paths, which has a and decoder. The encoder adopts masked self- many-to-one mapping to y since multiple CTC attention, which masks the future frames of a paths can correspond to the same text sequence. speech frame when encoding it and ensures For example, both the frame-level classification that each speech frame can only see its previ- outputs (CTC paths) “HHE;L;LOO” and ous frames to simulate the real-time streaming “;HHEEL;LO” are mapped to text sequence inputs.
Load more

Recommended publications
  • An Analysis of Sentence Segmentation Features For
    An Analysis of Sentence Segmentation Features for Broadcast News, Broadcast Conversations, and Meetings 12 1 Sebastien Cuendet1 Elizabeth Shriberg Benoit Favre [email protected] [email protected] [email protected] 1 James Fung1 Dilek Hakkani-Tur [email protected] [email protected] 2 1 ICSI SRI International 1947 Center Street 333 Ravenswood Ave Berkeley, CA 94704, USA Menlo Park, CA 94025, USA t sp eaking stylesbroadcast news BN broadcast ABSTRACT dieren conversations BC and facetoface multiparty meetings Information retrieval techniques for sp eech are based on MRDA We fo cus on the task of automatic sentence seg those develop ed for text and thus exp ect structured data as mentation or nding b oundaries of sentence units in the An essential task is to add sentence b oundary infor input otherwise unannotated devoid of punctuation capitaliza mation to the otherwise unannotated stream of words out tion or formatting stream of words output by a sp eech systems We analyze put by automatic sp eech recognition recognizer sentence segmentation p erformance as a function of feature Sentence segmentation is of particular imp ortance for sp eech ual versus automatic for news typ es and transcription man understanding applications b ecause techniques aimed at se sp eech meetings and a new corpus of broadcast conversa mantic pro cessing of sp eech inputsuch as machine trans tions Results show that overall features for broadcast lation question answering information extractionare typ news transfer well to meetings and broadcast
    [Show full text]
  • A Bayesian Framework for Word Segmentation: Exploring the Effects of Context
    Cognition 112 (2009) 21–54 Contents lists available at ScienceDirect Cognition journal homepage: www.elsevier.com/locate/COGNIT A Bayesian framework for word segmentation: Exploring the effects of context Sharon Goldwater a,*, Thomas L. Griffiths b, Mark Johnson c a School of Informatics, University of Edinburgh, Informatics Forum, 10 Crichton Street, Edinburgh, EH8 9AB, UK b Department of Psychology, University of California, Berkeley, CA, United States c Department of Cognitive and Linguistic Sciences, Brown University, United States article info abstract Article history: Since the experiments of Saffran et al. [Saffran, J., Aslin, R., & Newport, E. (1996). Statistical Received 30 May 2008 learning in 8-month-old infants. Science, 274, 1926–1928], there has been a great deal of Revised 11 March 2009 interest in the question of how statistical regularities in the speech stream might be used Accepted 13 March 2009 by infants to begin to identify individual words. In this work, we use computational mod- eling to explore the effects of different assumptions the learner might make regarding the nature of words – in particular, how these assumptions affect the kinds of words that are Keywords: segmented from a corpus of transcribed child-directed speech. We develop several models Computational modeling within a Bayesian ideal observer framework, and use them to examine the consequences of Bayesian Language acquisition assuming either that words are independent units, or units that help to predict other units. Word segmentation We show through empirical and theoretical results that the assumption of independence causes the learner to undersegment the corpus, with many two- and three-word sequences (e.g.
    [Show full text]
  • The Induction of Phonotactics for Speech Segmentation
    The Induction of Phonotactics for Speech Segmentation Converging evidence from computational and human learners Published by LOT phone: +31 30 253 6006 Trans 10 fax: +31 30 253 6406 3512 JK Utrecht e-mail: [email protected] The Netherlands http://www.lotschool.nl Cover illustration: Elbertus Majoor, Landschap-fantasie II (detail), 1958, gouache. ISBN: 978-94-6093-049-2 NUR 616 Copyright c 2011: Frans Adriaans. All rights reserved. The Induction of Phonotactics for Speech Segmentation Converging evidence from computational and human learners Het Induceren van Fonotactiek voor Spraaksegmentatie Convergerende evidentie uit computationele en menselijke leerders (met een samenvatting in het Nederlands) Proefschrift ter verkrijging van de graad van doctor aan de Universiteit Utrecht op gezag van de rector magnificus, prof.dr. J.C. Stoof, ingevolge het besluit van het college voor promoties in het openbaar te verdedigen op vrijdag 25 februari 2011 des middags te 2.30 uur door Frans Willem Adriaans geboren op 14 maart 1981 te Ooststellingwerf Promotor: Prof. dr. R.W.J. Kager To my parents, Ank & Piet CONTENTS acknowledgements xi 1 introduction 1 1.1 The speech segmentation problem ................. 1 1.2 Segmentation cues and their acquisition .............. 3 1.2.1 Theroleofsegmentationcuesinspokenwordrecognition 4 1.2.2 The acquisition of segmentation cues by infants ..... 7 1.3 The induction of phonotactics for speech segmentation ..... 11 1.3.1 Bottom-up versus top-down ................ 11 1.3.2 Computational modeling of phonotactic learning .... 12 1.3.3 Two hypotheses regarding the acquisition of phonotac- tics by infants ......................... 16 1.4 Learning mechanisms in early language acquisition ....... 17 1.4.1 Statistical learning .....................
    [Show full text]
  • Automatic Speech Segmentation
    Alaa Ehab Sakran et al, International Journal of Computer Science and Mobile Computing, Vol.6 Issue.4, April- 2017, pg. 308-315 Available Online at www.ijcsmc.com International Journal of Computer Science and Mobile Computing A Monthly Journal of Computer Science and Information Technology ISSN 2320–088X IMPACT FACTOR: 6.017 IJCSMC, Vol. 6, Issue. 4, April 2017, pg.308 – 315 A Review: Automatic Speech Segmentation Alaa Ehab Sakran 1, Sherif Mahdy Abdou 23, Salah Eldeen Hamid 4, Mohsen Rashwan 25 1 PhD. Program in Computer Science, Sudan University of Science and Technology, Khartoum, Sudan 2 Research & Development International (RDI®), Giza, Egypt 3 Department of IT, Faculty of Computers and Information, Cairo University, Giza, Egypt 4 Department of Engineering and Applied Sciences, Umm Al-Qura University, Mecca, Saudi Arabia 5 Department of Electronics and Communication Engineering, Cairo University, Giza, Egypt [email protected], [email protected], {sabdou, mrashwan} @rdi-eg.com I. INTRODUCTION Automated segmentation of speech signals has been under research for over 30 years. Many speech processing systems require segmentation of Speech waveform into principal acoustic units. Segmentation is a process of breaking down a speech signal into smaller units. Segmentation is the very primary step in any voiced activated systems like speech recognition systems and training of speech synthesis systems. Speech segmentation is performed utilizing Wavelet, Fuzzy methods, Artificial Neural Networks and Hidden Markov Model. [1] [2] In this research a review of basics of speech segmentation problem and state-of-the art solutions will be investigated. In next section main characteristics of speech signal will be discussed.
    [Show full text]
  • A Semi-Markov Model for Speech Segmentation with an Utterance-Break Prior
    A semi-Markov model for speech segmentation with an utterance-break prior Mark Sinclair1, Peter Bell1, Alexandra Birch2, Fergus McInnes1 1Centre for Speech Technology Research, 2Statistical Machine Translation Group School of Informatics, University of Edinburgh, Edinburgh EH8 9AB, UK fmark.sinclair, a.birch, peter.bell, [email protected] Abstract Previous work on detecting segmentation of sentence like Speech segmentation is the problem of finding the end points of units has looked at using features such as prosody [10], lan- a speech utterance for passing to an automatic speech recogni- guage model scores [11] [12], translation model scores [13] and tion (ASR) system. The quality of this segmentation can have syntactic constituents [14]. [15] presents a review of some of a large impact on the accuracy of the ASR system; in this pa- this work and also motivates tuning the segmentation of speech per we demonstrate that it can have an even larger impact on to the task at hand as we do in this paper. Our approach of mod- downstream natural language processing tasks – in this case, elling global sentence length distribution is orthogonal to much machine translation. We develop a novel semi-Markov model of this previous work, and combining these information sources which allows the segmentation of audio streams into speech ut- would be beneficial. There has been some previous work which terances which are optimised for the desired distribution of sen- attempts to exploit some of these cues [16]. However, in the tence lengths for the target domain. We compare this with exist- present paper, we have limited our focus to the use of statis- ing state-of-the-art methods and show that it is able to achieve tics of utterance durations and present a novel way of exploit- not only improved ASR performance, but also to yield signifi- ing this to select a globally optimal sequence from acoustically cant benefits to a speech translation task.
    [Show full text]
  • Seminar, Tue 4:15Pm
    On an exceptionally hot evening early in July a young man came out of the garret in which he lodged in S. Place and walked slowly, as though in hesitation, towards K. bridge. He had successfully avoided meeting his landlady on the staircase. His garret was under the roof of a high, five-storied house and was more like a cupboard than a room. The landlady who provided him with garret, dinners, and attendance, lived on the floor below, and every time he went out he was obliged to pass her kitchen, the door of which invariably stood open. And each time he passed, the young man had a sick, frightened feeling, which made him scowl and feel ashamed. He was hopelessly in debt to his landlady, and was afraid of meeting her. This was not because he was cowardly and abject, quite the contrary; but for some time past he had been in an overstrained irritable condition, verging on hypochondria. He had become so completely absorbed in himself, and isolated from his fellows that he dreaded meeting, not only his landlady, but anyone at all. He was crushed by poverty, but the anxieties of his position had ofNatural late ceased to weigh upon him.language He had given up attending to matters of practical importance; he had lost all desire to do so. Nothing that any landlady could do had a real terror for him. But to be stopped on the stairs, to be forced to listen to her trivial, irrelevant gossip, to pestering demands for payment, threats and complaints, and to rack his brains for excuses, to prevaricate, to lie—no, rather than that, he would creep down the stairs like a cat andprocessing slip out unseen.
    [Show full text]
  • The Polysemy of the Words That Children Learn Over Time Arxiv
    The Polysemy of the Words that Children Learn over Time Bernardino Casas [email protected] 1 Neus Català [email protected] 2 Ramon Ferrer-i-Cancho [email protected] 1 Antoni Hernández-Fernández [email protected] 1,3 Jaume Baixeries [email protected] 1 1. Complexity & Quantitative Linguistics Lab, Departament de Ciències de la Computació, Laboratory for Relational Algorithmics, Complexity and Learning (LARCA), Universitat Politècnica de Catalunya, Barcelona, Catalonia. 2. Complexity & Quantitative Linguistics Lab, Departament de Ciències de la Computació, Center for Language and Speech Technologies and Applications (TALP Research Center), Universitat Politècnica de Catalunya, Barcelona, Catalonia. 3. Institut de Ciències de l0Educació, Universitat Politècnica de Catalunya, Barcelona, Catalonia. arXiv:1611.08807v2 [cs.CL] 26 Mar 2019 1 Abstract Here we study polysemy as a potential learning bias in vocabulary learning in children. Words of low polysemy could be preferred as they reduce the disambiguation effort for the listener. However, such preference could be a side-effect of another bias: the preference of chil- dren for nouns in combination with the lower polysemy of nouns with respect to other part-of-speech categories. Our results show that mean polysemy in children increases over time in two phases, i.e. a fast growth till the 31st month followed by a slower tendency towards adult speech. In contrast, this evolution is not found in adults interacting with children. This suggests that children have a preference for non-polysemous words in their early stages of vocabulary acquisition. Interestingly, the evolutionary pat- tern described above weakens when controlling for syntactic category (noun, verb, adjective or adverb) but it does not disappear completely, suggesting that it could result from a combination of a standalone bias for low polysemy and a preference for nouns.
    [Show full text]
  • AUTOMATIC LINGUISTIC SEGMENTATION of CONVERSATIONAL SPEECH Andreas Stolcke Elizabeth Shriberg
    AUTOMATIC LINGUISTIC SEGMENTATION OF CONVERSATIONAL SPEECH Andreas Stolcke Elizabeth Shriberg Speech Technology and Research Laboratory SRI International, Menlo Park, CA 94025 [email protected] [email protected] ABSTRACT could use the knowledge incorporated in an automatic segmenter to help end-point a user’s speech input. A speech indexing and re- As speech recognition moves toward more unconstrained domains trieval system (such as for transcribed broadcast audio) could pro- such as conversational speech, we encounter a need to be able to cess its data in more meaningful units if the locations of linguistic segment (or resegment) waveforms and recognizer output into lin- segment boundaries were known. guistically meaningful units, such a sentences. Toward this end, we present a simple automatic segmenter of transcripts based on Our main motivation for the work reported here comes from speech N-gram language modeling. We also study the relevance of sev- language modeling. Experiments at the 1995 Johns Hopkins Lan- eral word-level features for segmentation performance. Using only guage Modeling Workshop showed that the quality of a language word-level information, we achieve 85% recall and 70% precision model (LM) can be improved if both training and test data are seg- on linguistic boundary detection. mented linguistically, rather than acoustically [8]. We showed in [10] and [9] that proper modeling of filled pauses requires knowl- 1. INTRODUCTION edge of linguistic segment boundaries. We found for example that Today’s large-vocabulary speech recognizers typically prefer to pro- segment-internal filled pauses condition the following words quite cess a few tens of seconds of speech at a time, to keep the time and differently from segment-initial filled pauses.
    [Show full text]
  • Segmentation and Morphology
    Segmentation and morphology John A. Goldsmith1 Departments of Linguistics and Computer Science, The University of Chicago [email protected] A draft chapter for the Blackwell Computational Linguistics and Natural Lan- guage Processing Handbook, edited by Alex Clark, Chris Fox and Shalom Lappin. This draft formatted on 2nd May 2009. Page: 1 job: goldsmith macro: handbook.cls date/time: 2-May-2009/17:28 Page: 2 job: goldsmith macro: handbook.cls date/time: 2-May-2009/17:28 Contents Segmentation and morphology John A. Goldsmith ............................................... 1 1 Introduction................................................. 4 1.1 Generalremarks......................................... 4 1.2 Morphology............................................. 6 1.3 Staticanddynamicmetaphors ............................ 9 2 Unsupervisedlearningofwords ................................ 11 2.1 Thetwoproblemsofwordsegmentation.................... 11 2.2 Trawlingforchunks...................................... 14 2.3 Word Boundary detectors . 21 2.4 Successesandfailuresinwordsegmentation................. 22 3 Unsupervisedlearningofmorphology............................ 23 3.1 ZelligHarris ............................................ 23 3.2 Usingdescriptionlength.................................. 24 3.3 Workinthefield ........................................ 27 4 Implementingcomputationalmorphologies....................... 29 4.1 Finitestatetransducers................................... 30 4.2 Morphophonology.......................................
    [Show full text]
  • Unsupervised Segmentation of Audio Speech Using the Voting Experts Algorithm Matthew Im Ller Adam Miller Iowa State University
    Iowa State University Capstones, Theses and Graduate Theses and Dissertations Dissertations 2009 Unsupervised Segmentation of Audio Speech Using the Voting Experts Algorithm Matthew iM ller Adam Miller Iowa State University Follow this and additional works at: https://lib.dr.iastate.edu/etd Part of the Computer Sciences Commons Recommended Citation Miller, Matthew Miller Adam, "Unsupervised Segmentation of Audio Speech Using the Voting Experts Algorithm" (2009). Graduate Theses and Dissertations. 10725. https://lib.dr.iastate.edu/etd/10725 This Thesis is brought to you for free and open access by the Iowa State University Capstones, Theses and Dissertations at Iowa State University Digital Repository. It has been accepted for inclusion in Graduate Theses and Dissertations by an authorized administrator of Iowa State University Digital Repository. For more information, please contact [email protected]. Unsupervised segmentation of audio speech using the Voting Experts algorithm by Matthew Adam Miller A thesis submitted to the graduate faculty in partial ful¯llment of the requirements for the degree of MASTER OF SCIENCE Major: Computer Science Program of Study Committee: Alexander Stoytchev, Major Professor Giora Slutzki Pavan Aduri Iowa State University Ames, Iowa 2009 Copyright °c Matthew Adam Miller, 2009. All rights reserved. ii DEDICATION To Cat. iii TABLE OF CONTENTS LIST OF TABLES . vi LIST OF FIGURES . viii ACKNOWLEDGEMENTS . xi ABSTRACT . xii CHAPTER 1. INTRODUCTION . 1 1.1 Research Questions . 4 CHAPTER 2. REVIEW OF LITERATURE . 6 2.1 Psychology of Speech Segmentation . 7 2.1.1 Statistical Learning . 8 2.1.2 Isolated Words . 10 2.1.3 Allophonic Variation . 11 2.1.4 Metrical Cues .
    [Show full text]
  • The Universal Dependencies Treebank of Spoken Slovenian
    The Universal Dependencies Treebank of Spoken Slovenian Kaja Dobrovoljc1, Joakim Nivre2 1Institute for Applied Slovene Studies Trojina, Ljubljana, Slovenia 1Department of Slovenian Studies, Faculty of Arts, University of Ljubljana 2Department of Linguistics and Philology, Uppsala University [email protected], joakim.nivre@lingfil.uu.se Abstract This paper presents the construction of an open-source dependency treebank of spoken Slovenian, the first syntactically annotated collection of spontaneous speech in Slovenian. The treebank has been manually annotated using the Universal Dependencies annotation scheme, a one-layer syntactic annotation scheme with a high degree of cross-modality, cross-framework and cross-language interoper- ability. In this original application of the scheme to spoken language transcripts, we address a wide spectrum of syntactic particularities in speech, either by extending the scope of application of existing universal labels or by proposing new speech-specific extensions. The initial analysis of the resulting treebank and its comparison with the written Slovenian UD treebank confirms significant syntactic differences between the two language modalities, with spoken data consisting of shorter and more elliptic sentences, less and simpler nominal phrases, and more relations marking disfluencies, interaction, deixis and modality. Keywords: dependency treebank, spontaneous speech, Universal Dependencies 1. Introduction actually out-performs state-of-the-art pipeline approaches (Rasooli and Tetreault, 2013; Honnibal and Johnson, 2014). It is nowadays a well-established fact that data-driven pars- Such heterogeneity of spoken language annotation schemes ing systems used in different speech-processing applica- inevitably leads to a restricted usage of existing spoken tions benefit from learning on annotated spoken data, rather language treebanks in linguistic research and parsing sys- than using models built on written language observation.
    [Show full text]
  • Sentence Boundary Detection Based on Parallel Lexical and Acoustic Models
    INTERSPEECH 2016 September 8–12, 2016, San Francisco, USA Sentence Boundary Detection Based on Parallel Lexical and Acoustic Models Xiaoyin Che, Sheng Luo, Haojin Yang, Christoph Meinel Hasso Plattner Institute, University of Potsdam Prof.-Dr.-Helmert-Str. 2-3, 14482, Potsdam, Germany fxiaoyin.che, sheng.luo, haojin.yang, [email protected] Abstract models on ASR-available TED Talks and run experiments about In this paper we propose a solution that detects sentence bound- our 2-stage joint decision scheme with different combination of ary from speech transcript. First we train a pure lexical model models. In the end comes the conclusion. with deep neural network, which takes word vectors as the only input feature. Then a simple acoustic model is also prepared. 2. Multi-Modal Structure Analysis Because the models work independently, they can be trained with different data. In next step, the posterior probabilities of The structures of multi-modal solutions can be different and both lexical and acoustic models will be involved in a heuristic have been discussed before [12]. Some researchers propose a 2-stage joint decision scheme to classify the sentence boundary single hybrid model, which takes all possible features, no matter positions. This approach ensures that the models can be up- lexical or acoustic, together as the model input [13, 14, 15, 16], dated or switched freely in actual use. Evaluation on TED Talks as shown in Figure 1-a. Some others apply a structure of shows that the proposed lexical model can achieve good results: sequential models, in which the output of model A is fed 75.5% accuracy on error-involved ASR transcripts and 82.4% into model B, as shown in Figure 1-b, where model A ac- on error-free manual references.
    [Show full text]