DOCSLIB.ORG

Disambiguating Heteronyms in Speech Synthesis

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Disambiguating Heteronyms in Speech Synthesis (19) TZZ¥Z¥ ¥ _T (11) EP 3 032 532 A1 (12) EUROPEAN PATENT APPLICATION (43) Date of publication: (51) Int Cl.: 15.06.2016 Bulletin 2016/24 G10L 13/08 (2013.01) G10L 15/22 (2006.01) (21) Application number: 15196748.6 (22) Date of filing: 27.11.2015 (84) Designated Contracting States: (71) Applicant: APPLE INC. AL AT BE BG CH CY CZ DE DK EE ES FI FR GB Cupertino, CA 95014 (US) GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR (72) Inventors: Designated Extension States: • HENTON, Caroline BA ME Cupertino, CA California 95014 (US) Designated Validation States: • NAIK, Devang MA MD Cupertino, CA California 95014 (US) (30) Priority: 09.12.2014 US 201462089464 P (74) Representative: Barnfather, Karl Jon 12.12.2014 US 201414569517 Withers & Rogers LLP 4 More London Riverside London SE1 2AU (GB) (54) DISAMBIGUATING HETERONYMS IN SPEECH SYNTHESIS (57) Systems and processes for disambiguating het- mined based on at least one of the phonemic string or eronyms in speech synthesis are provided. In one exam- usingan n-gram languagemodel of the automatic speech ple process, a speech input containing a heteronym can recognition system. A dialogue response to the speech be received from a user. The speech input can be proc- input can be generated where the dialogue response can essed using an automatic speech recognition system to include the heteronym. The dialogue response can be determine a phonemic string corresponding to the het- outputted as a speech output. The heteronym in the di- eronym as pronounced by the user in the speech input. alogue response can be pronounced in the speech output A correct pronunciation of the heteronym can be deter- according to the correct pronunciation. EP 3 032 532 A1 Printed by Jouve, 75001 PARIS (FR) 1 EP 3 032 532 A1 2 Description ing to the heteronym as pronounced by the user in the speech input and a frequency of occurrence of an n-gram Cross-Reference to Related Application with respect to a corpus. The n-gram can include the heteronym and the one or more additional words and the [0001] This application claims priority from U.S. Provi- 5 heteronym in the n-gram can be associated with a first sional Serial No. 62/089,464, filed on December 9, 2014, pronunciation. A correct pronunciation of the heteronym entitled "Disambiguating Heteronyms in Speech Synthe- can be based on at least one of the phonemic string and sis," and U.S. Non-Provisional Serial No. 14/569,517, the frequency of occurrence of the n-gram. A dialogue filed on December 12, 2014, entitled "Disambiguating response to the speech input can be generated where Heteronyms in Speech Synthesis," which are hereby in- 10 the dialogue response can include the heteronym. The corporated by reference in their entirety for all purposes. dialogue response can be output as a speech output. The heteronym in the dialogue response can be pro- Field nounced in the speech output according to the deter- mined correct pronunciation. [0002] This relates generally to digital assistants and, 15 more specifically, to disambiguating heteronyms in Brief Description of the Drawings speech synthesis. [0006] Background 20 FIG. 1 illustrates a system and environment for im- [0003] Intelligent automated assistants (or digital as- plementing a digital assistant according to various sistants) can provide a beneficial interface between hu- examples. man users and electronic devices. Such assistants can allow users to interact with devices or systems using nat- FIG. 2 illustrates a user device implementing the cli- ural language in spoken and/or text forms. For example, 25 ent-side portion of a digital assistant according to a user can provide a speech input to a digital assistant various examples. associated with the electronic device. The digital assist- ant can interpret the user’s intent from the speech input FIG. 3A illustrates a digital assistant system or a and operationalize the user’s intent into tasks. The tasks server portion thereof according to various exam- can then be performed by executing one or more services 30 ples. of the electronic device and a relevant speech output can be returned to the user in natural language form. FIG. 3B illustrates the functions of the digital assist- [0004] Occasionally, speech outputs generated by dig- ant shown in FIG. 3A according to various examples. ital assistants can contain heteronyms. A heteronym can be each of two or more words that are spelled identically 35 FIG. 3C illustrates a portion of an ontology according but have different pronunciations and meanings. For ex- to various examples. ample, a user can provide a speech input to a digital assistant requesting the weather in Nice, France. The FIG. 4 illustrates a process for operating a digital digital assistant can return a relevant speech output such assistant according to various examples. as, "Here is the weather in Nice, France." In this example, 40 the speech output contains the heteronym "nice," which FIG. 5 illustrates a functional block diagram of an can have one pronunciation as a correct noun and a dif- electronic device according to various examples. ferent pronunciation as an adjective. Conventionally, dig- ital assistants can have difficult disambiguating hetero- Detailed Description nyms and thus speech outputs containing heteronyms 45 can often be pronounced incorrectly. This can result in a [0007] In the following description of examples, refer- poor user experience in interacting with the digital assist- ence is made to the accompanying drawings in which it ant. is shown by way of illustration specific examples that can be practiced. It is to be understood that other examples Summary 50 can be used and structural changes can be made without departing from the scope of the various examples. [0005] Systems and processes for disambiguating het- [0008] Systems and processes for disambiguating het- eronyms in speech synthesis are provided. In an example eronyms in speech synthesis are provided. In one exam- process, a speech input can be received from a user. ple process, a speech input containing a heteronym can The speech input can contain a heteronym and one or 55 be received from a user. Contextual data associated with more additional words. The speech input can be proc- the speech input can be received. The speech input can essed using an automatic speech recognition system to be processed using an automatic speech recognition determine at least one of a phonemic string correspond- system to determine a text string corresponding to the 2 3 EP 3 032 532 A1 4 speech input. Based on the text string, an actionable in- west gate." The user can also request the performance tent can be determined using a natural language proc- of a task, for example, "Please invite my friends to my essor. A dialogue response to the speech input can be girlfriend’s birthday party next week." In response, the generated where the dialogue response can include the digital assistant can acknowledge the request by saying heteronym. A correct pronunciation of the heteronym can 5 "Yes, right away," and then send a suitable calendar in- be determined using an n-gram language model of the vite on behalf of the user to each of the user’s friends automatic speech recognition system or based on at least listed in the user’s electronic address book. During per- one of the speech input, the actionable intent, and the formance of a requested task, the digital assistant can contextual data. The dialogue response can be output sometimes interact with the user in a continuous dialogue as a speech output and the heteronym can be pro-10 involving multiple exchanges of information over an ex- nounced in the speech output according to the correct tended period of time. There are numerous other ways pronunciation. of interacting with a digital assistant to request informa- [0009] By utilizing at least one of the speech input, the tion or performance of various tasks. In addition to pro- n-gram language model, the actionable intent, and the viding verbal responses and taking programmed actions, contextual data as a knowledge source for disambigua- 15 the digital assistant can also provide responses in other ting heteronyms, the pronunciation of heteronyms in the visual or audio forms, e.g., as text, alerts, music, videos, speech output can be synthesized more accurately, animations, etc. thereby improving user experience. Further, leveraging [0012] As shown in FIG. 1, in some examples, a digital the automatic speech recognition system and the natural assistant can be implemented according to a client-serv- language processorto disambiguate heteronyms can ob- 20 er model. The digital assistant can include a client-side viate the need to implement additional resources in the portion 102a, 102b (hereafter "DA client 102") executed speech synthesizer for the same purpose. For example, on a user device 104a, 104b, and a server-side portion additional language models need not be implemented in 106 (hereafter "DA server 106") executed on a server the speech synthesizer to disambiguate heteronyms. system 108. The DA client 102 can communicate with This enables digital assistants to operate with greater 25 the DA server 106 through one or more networks 110. efficiency and fewer resources. The DA client 102 can provide client-side functionalities such as user-facing input and output processing and 1. System and Environment communication with the DA-server 106. The DA server 106 can provide server-side functionalities for any [0010] FIG. 1 illustrates a block diagram of a system 30 number of DA-clients 102 each residing on a respective 100 according to various examples.
Load more

Recommended publications
  • The Role of Higher-Level Linguistic Features in HMM-Based Speech Synthesis
    INTERSPEECH 2010 The role of higher-level linguistic features in HMM-based speech synthesis Oliver Watts, Junichi Yamagishi, Simon King Centre for Speech Technology Research, University of Edinburgh, UK [email protected] [email protected] [email protected] Abstract the annotation of a synthesiser’s training data on listeners’ re- action to the speech produced by that synthesiser is still unclear We analyse the contribution of higher-level elements of the lin- to us. In particular, we suspect that features such as pitch ac- guistic specification of a data-driven speech synthesiser to the cent type which might be useful in voice-building if labelled naturalness of the synthetic speech which it generates. The reliably, are under-used in conventional systems because of la- system is trained using various subsets of the full feature-set, belling/prediction errors. For building the systems to be evalu- in which features relating to syntactic category, intonational ated we therefore used data for which hand labelled annotation phrase boundary, pitch accent and boundary tones are selec- of ToBI events is available. This allows us to compare ideal sys- tively removed. Utterances synthesised by the different config- tems built from a corpus where these higher-level features are urations of the system are then compared in a subjective evalu- accurately annotated with more conventional systems that rely ation of their naturalness. exclusively on prediction of these features from text for annota- The work presented forms background analysis for an on- tion of their training data. going set of experiments in performing text-to-speech (TTS) conversion based on shallow features: features that can be triv- 2.
    [Show full text]
  • The RACAI Text-To-Speech Synthesis System
    The RACAI Text-to-Speech Synthesis System Tiberiu Boroș, Radu Ion, Ștefan Daniel Dumitrescu Research Institute for Artificial Intelligence “Mihai Drăgănescu”, Romanian Academy (RACAI) [email protected], [email protected], [email protected] of standalone Natural Language Processing (NLP) Tools Abstract aimed at enabling text-to-speech (TTS) synthesis for less- This paper describes the RACAI Text-to-Speech (TTS) entry resourced languages. A good example is the case of for the Blizzard Challenge 2013. The development of the Romanian, a language which poses a lot of challenges for TTS RACAI TTS started during the Metanet4U project and the synthesis mainly because of its rich morphology and its system is currently part of the METASHARE platform. This reduced support in terms of freely available resources. Initially paper describes the work carried out for preparing the RACAI all the tools were standalone, but their design allowed their entry during the Blizzard Challenge 2013 and provides a integration into a single package and by adding a unit selection detailed description of our system and future development speech synthesis module based on the Pitch Synchronous directions. Overlap-Add (PSOLA) algorithm we were able to create a fully independent text-to-speech synthesis system that we refer Index Terms: speech synthesis, unit selection, concatenative to as RACAI TTS. 1. Introduction A considerable progress has been made since the start of the project, but our system still requires development and Text-to-speech (TTS) synthesis is a complex process that although considered premature, our participation in the addresses the task of converting arbitrary text into voice.
    [Show full text]
  • Synthesis and Recognition of Speech Creating and Listening to Speech
    ISSN 1883-1974 (Print) ISSN 1884-0787 (Online) National Institute of Informatics News NII Interview 51 A Combination of Speech Synthesis and Speech Oct. 2014 Recognition Creates an Affluent Society NII Special 1 “Statistical Speech Synthesis” Technology with a Rapidly Growing Application Area NII Special 2 Finding Practical Application for Speech Recognition Feature Synthesis and Recognition of Speech Creating and Listening to Speech A digital book version of “NII Today” is now available. http://www.nii.ac.jp/about/publication/today/ This English language edition NII Today corresponds to No. 65 of the Japanese edition [Advance Notice] Great news! NII Interview Yamagishi-sensei will create my voice! Yamagishi One result is a speech translation sys- sound. Bit (NII Character) A Combination of Speech Synthesis tem. This system recognizes speech and translates it Ohkawara I would like your comments on the fu- using machine translation to synthesize speech, also ture challenges. and Speech Recognition automatically translating it into every language to Ono The challenge for speech recognition is how speak. Moreover, the speech is created with a human close it will come to humans in the distant speech A Word from the Interviewer voice. In second language learning, you can under- case. If this study is advanced, it will be possible to Creates an Affluent Society stand how you should pronounce it with your own summarize the contents of a meeting and to automati- voice. If this is further advanced, the system could cally take the minutes. If a robot understands the con- have an actor in a movie speak in a different language tents of conversations by multiple people in a natural More and more people have begun to use smart- a smartphone, I use speech input more often.
    [Show full text]
  • The Role of Speech Processing in Human-Computer Intelligent Communication
    THE ROLE OF SPEECH PROCESSING IN HUMAN-COMPUTER INTELLIGENT COMMUNICATION Candace Kamm, Marilyn Walker and Lawrence Rabiner Speech and Image Processing Services Research Laboratory AT&T Labs-Research, Florham Park, NJ 07932 Abstract: We are currently in the midst of a revolution in communications that promises to provide ubiquitous access to multimedia communication services. In order to succeed, this revolution demands seamless, easy-to-use, high quality interfaces to support broadband communication between people and machines. In this paper we argue that spoken language interfaces (SLIs) are essential to making this vision a reality. We discuss potential applications of SLIs, the technologies underlying them, the principles we have developed for designing them, and key areas for future research in both spoken language processing and human computer interfaces. 1. Introduction Around the turn of the twentieth century, it became clear to key people in the Bell System that the concept of Universal Service was rapidly becoming technologically feasible, i.e., the dream of automatically connecting any telephone user to any other telephone user, without the need for operator assistance, became the vision for the future of telecommunications. Of course a number of very hard technical problems had to be solved before the vision could become reality, but by the end of 1915 the first automatic transcontinental telephone call was successfully completed, and within a very few years the dream of Universal Service became a reality in the United States. We are now in the midst of another revolution in communications, one which holds the promise of providing ubiquitous service in multimedia communications.
    [Show full text]
  • Voice User Interface on the Web Human Computer Interaction Fulvio Corno, Luigi De Russis Academic Year 2019/2020 How to Create a VUI on the Web?
    Voice User Interface On The Web Human Computer Interaction Fulvio Corno, Luigi De Russis Academic Year 2019/2020 How to create a VUI on the Web? § Three (main) steps, typically: o Speech Recognition o Text manipulation (e.g., Natural Language Processing) o Speech Synthesis § We are going to start from a simple application to reach a quite complex scenario o by using HTML5, JS, and PHP § Reminder: we are interested in creating an interactive prototype, at the end 2 Human Computer Interaction Weather Web App A VUI for "chatting" about the weather Base implementation at https://github.com/polito-hci-2019/vui-example 3 Human Computer Interaction Speech Recognition and Synthesis § Web Speech API o currently a draft, experimental, unofficial HTML5 API (!) o https://wicg.github.io/speech-api/ § Covers both speech recognition and synthesis o different degrees of support by browsers 4 Human Computer Interaction Web Speech API: Speech Recognition § Accessed via the SpeechRecognition interface o provides the ability to recogniZe voice from an audio input o normally via the device's default speech recognition service § Generally, the interface's constructor is used to create a new SpeechRecognition object § The SpeechGrammar interface can be used to represent a particular set of grammar that your app should recogniZe o Grammar is defined using JSpeech Grammar Format (JSGF) 5 Human Computer Interaction Speech Recognition: A Minimal Example const recognition = new window.SpeechRecognition(); recognition.onresult = (event) => { const speechToText = event.results[0][0].transcript;
    [Show full text]
  • A Multimodal User Interface for an Assistive Robotic Shopping Cart
    electronics Article A Multimodal User Interface for an Assistive Robotic Shopping Cart Dmitry Ryumin 1 , Ildar Kagirov 1,* , Alexandr Axyonov 1, Nikita Pavlyuk 1, Anton Saveliev 1, Irina Kipyatkova 1, Milos Zelezny 2, Iosif Mporas 3 and Alexey Karpov 1,* 1 St. Petersburg Institute for Informatics and Automation of the Russian Academy of Sciences (SPIIRAS), St. Petersburg Federal Research Center of the Russian Academy of Sciences (SPC RAS), 199178 St. Petersburg, Russia; [email protected] (D.R.); [email protected] (A.A.); [email protected] (N.P.); [email protected] (A.S.); [email protected] (I.K.) 2 Department of Cybernetics, Faculty of Applied Sciences, University of West Bohemia, 301 00 Pilsen, Czech Republic; [email protected] 3 School of Engineering and Computer Science, University of Hertfordshire, Hatfield, Herts AL10 9AB, UK; [email protected] * Correspondence: [email protected] (I.K.); [email protected] (A.K.) Received: 2 November 2020; Accepted: 5 December 2020; Published: 8 December 2020 Abstract: This paper presents the research and development of the prototype of the assistive mobile information robot (AMIR). The main features of the presented prototype are voice and gesture-based interfaces with Russian speech and sign language recognition and synthesis techniques and a high degree of robot autonomy. AMIR prototype’s aim is to be used as a robotic cart for shopping in grocery stores and/or supermarkets. Among the main topics covered in this paper are the presentation of the interface (three modalities), the single-handed gesture recognition system (based on a collected database of Russian sign language elements), as well as the technical description of the robotic platform (architecture, navigation algorithm).
    [Show full text]
  • Voice Assistants and Smart Speakers in Everyday Life and in Education
    Informatics in Education, 2020, Vol. 19, No. 3, 473–490 473 © 2020 Vilnius University, ETH Zürich DOI: 10.15388/infedu.2020.21 Voice Assistants and Smart Speakers in Everyday Life and in Education George TERZOPOULOS, Maya SATRATZEMI Department of Applied Informatics, University of Macedonia, Thessaloniki, Greece Email: [email protected], [email protected] Received: November 2019 Abstract. In recent years, Artificial Intelligence (AI) has shown significant progress and its -po tential is growing. An application area of AI is Natural Language Processing (NLP). Voice as- sistants incorporate AI by using cloud computing and can communicate with the users in natural language. Voice assistants are easy to use and thus there are millions of devices that incorporates them in households nowadays. Most common devices with voice assistants are smart speakers and they have just started to be used in schools and universities. The purpose of this paper is to study how voice assistants and smart speakers are used in everyday life and whether there is potential in order for them to be used for educational purposes. Keywords: artificial intelligence, smart speakers, voice assistants, education. 1. Introduction Emerging technologies like virtual reality, augmented reality and voice interaction are reshaping the way people engage with the world and transforming digital experiences. Voice control is the next evolution of human-machine interaction, thanks to advances in cloud computing, Artificial Intelligence (AI) and the Internet of Things (IoT). In the last years, the heavy use of smartphones led to the appearance of voice assistants such as Apple’s Siri, Google’s Assistant, Microsoft’s Cortana and Amazon’s Alexa.
    [Show full text]
  • Models of Speech Synthesis ROLF CARLSON Department of Speech Communication and Music Acoustics, Royal Institute of Technology, S-100 44 Stockholm, Sweden
    Proc. Natl. Acad. Sci. USA Vol. 92, pp. 9932-9937, October 1995 Colloquium Paper This paper was presented at a colloquium entitled "Human-Machine Communication by Voice," organized by Lawrence R. Rabiner, held by the National Academy of Sciences at The Arnold and Mabel Beckman Center in Irvine, CA, February 8-9,1993. Models of speech synthesis ROLF CARLSON Department of Speech Communication and Music Acoustics, Royal Institute of Technology, S-100 44 Stockholm, Sweden ABSTRACT The term "speech synthesis" has been used need large amounts of speech data. Models working close to for diverse technical approaches. In this paper, some of the the waveform are now typically making use of increased unit approaches used to generate synthetic speech in a text-to- sizes while still modeling prosody by rule. In the middle of the speech system are reviewed, and some of the basic motivations scale, "formant synthesis" is moving toward the articulatory for choosing one method over another are discussed. It is models by looking for "higher-level parameters" or to larger important to keep in mind, however, that speech synthesis prestored units. Articulatory synthesis, hampered by lack of models are needed not just for speech generation but to help data, still has some way to go but is yielding improved quality, us understand how speech is created, or even how articulation due mostly to advanced analysis-synthesis techniques. can explain language structure. General issues such as the synthesis of different voices, accents, and multiple languages Flexibility and Technical Dimensions are discussed as special challenges facing the speech synthesis community.
    [Show full text]
  • Attention, I'm Trying to Speak Cs224n Project: Speech Synthesis
    Attention, I’m Trying to Speak CS224n Project: Speech Synthesis Akash Mahajan Management Science & Engineering Stanford University [email protected] Abstract We implement an end-to-end parametric text-to-speech synthesis model that pro- duces audio from a sequence of input characters, and demonstrate that it is possi- ble to build a convolutional sequence to sequence model with reasonably natural voice and pronunciation from scratch in well under $75. We observe training the attention to be a bottleneck and experiment with 2 modifications. We also note interesting model behavior and insights during our training process. Code for this project is available on: https://github.com/akashmjn/cs224n-gpu-that-talks. 1 Introduction We have come a long way from the ominous robotic sounding voices used in the Radiohead classic 1. If we are to build significant voice interfaces, we need to also have good feedback that can com- municate with us clearly, that can be setup cheaply for different languages and dialects. There has recently also been an increased interest in generative models for audio [6] that could have applica- tions beyond speech to music and other signal-like data. As discussed in [13] it is fascinating that is is possible at all for an end-to-end model to convert a highly ”compressed” source - text - into a substantially more ”decompressed” form - audio. This is a testament to the sheer power and flexibility of deep learning models, and has been an interesting and surprising insight in itself. Recently, results from Tachibana et. al. [11] reportedly produce reasonable-quality speech without requiring as large computational resources as Tacotron [13] and Wavenet [7].
    [Show full text]
  • Quantifying the Effects of Prosody Modulation on User Engagement
    Quantifying the Effects of Prosody Modulation onUser Engagement and Satisfaction in Conversational Systems Jason Ingyu Choi Eugene Agichtein Computer Science Department Computer Science Department Emory University Emory University [email protected] [email protected] ABSTRACT As voice-based assistants such as Alexa, Siri, and Google Assistant become ubiquitous, users increasingly expect to maintain natural and informative conversations with such systems. However, for an open-domain conversational system to be coherent and engag- ing, it must be able to maintain the user’s interest for extended periods, without sounding “boring” or “annoying”. In this paper, we investigate one natural approach to this problem, of modulat- ing response prosody, i.e., changing the pitch and cadence of the response to indicate delight, sadness or other common emotions, as well as using pre-recorded interjections. Intuitively, this approach should improve the naturalness of the conversation, but attempts to quantify the effects of prosodic modulation on user satisfaction and engagement remain challenging. To accomplish this, we report results obtained from a large-scale empirical study that measures the effects of prosodic modulation on user behavior and engage- ment across multiple conversation domains, both immediately after each turn, and at the overall conversation level. Our results indicate that the prosody modulation significantly increases both immedi- Figure 1: Sample human-machine conversation from our ate and overall user satisfaction. However, since the effects vary system. The red texts show response examples after insert- across different domains, we verify that prosody modulations do ing prerecorded Speechcons to convey artificial emotion. not substitute for coherent, informative content of the responses.
    [Show full text]
  • Lecture 5: Part-Of-Speech Tagging
    CS447: Natural Language Processing http://courses.engr.illinois.edu/cs447 Lecture 5: Part-of-Speech Tagging Julia Hockenmaier [email protected] 3324 Siebel Center POS tagging Tagset: NNP: proper noun CD: numeral, POS tagger JJ: adjective, ... Raw text Tagged text Pierre_NNP Vinken_NNP ,_, 61_CD Pierre Vinken , 61 years old years_NNS old_JJ ,_, will_MD join_VB , will join the board as a the_DT board_NN as_IN a_DT nonexecutive director Nov. nonexecutive_JJ director_NN Nov._NNP 29 . 29_CD ._. CS447: Natural Language Processing (J. Hockenmaier) !2 Why POS tagging? POS tagging is a prerequisite for further analysis: –Speech synthesis: How to pronounce “lead”? INsult or inSULT, OBject or obJECT, OVERflow or overFLOW," DIScount or disCOUNT, CONtent or conTENT –Parsing: What words are in the sentence? –Information extraction: Finding names, relations, etc. –Machine Translation: The noun “content” may have a different translation from the adjective. CS447: Natural Language Processing (J. Hockenmaier) !3 POS Tagging Words often have more than one POS: " -The back door (adjective) -On my back (noun) -Win the voters back (particle) -Promised to back the bill (verb)" The POS tagging task is to determine the POS tag " for a particular instance of a word. " Since there is ambiguity, we cannot simply look up the correct POS in a dictionary. These examples from Dekang Lin CS447: Natural Language Processing (J. Hockenmaier) !4 Defining a tagset CS447: Natural Language Processing (J. Hockenmaier) !5 Defining a tag set We have to define an inventory of labels for the word classes (i.e. the tag set)" -Most taggers rely on models that have to be trained on annotated (tagged) corpora.
    [Show full text]
  • Investigation of Using Continuous Representation of Various Linguistic Units in Neural Network Based Text-To-Speech Synthesis
    IEICE TRANS. INF. & SYST., VOL.E99–D, NO.10 OCTOBER 2016 2471 PAPER Special Section on Recent Advances in Machine Learning for Spoken Language Processing Investigation of Using Continuous Representation of Various Linguistic Units in Neural Network Based Text-to-Speech Synthesis Xin WANG†,††a), Shinji TAKAKI†, Nonmembers, and Junichi YAMAGISHI†,††,†††, Member SUMMARY Building high-quality text-to-speech (TTS) systems with- tated. This results in very accurate and high-quality syn- out expert knowledge of the target language and/or time-consuming man- thetic speech. However, it is laborious to collect such the ual annotation of speech and text data is an important yet challenging re- databases. This becomes the major barrier, especially when search topic. In this kind of TTS system, it is vital to find representation of the input text that is both effective and easy to acquire. Recently, the we need to build a TTS system in a new language in which continuous representation of raw word inputs, called “word embedding”, speech and linguistic resources are lacking. Even for the has been successfully used in various natural language processing tasks. It major languages such as English, we encounter a similar has also been used as the additional or alternative linguistic input features problem when we scale up the size of the database or switch to a neural-network-based acoustic model for TTS systems. In this pa- per, we further investigate the use of this embedding technique to represent to a new domain. phonemes, syllables and phrases for the acoustic model based on the re- Recently, the continuous representation of raw word in- current and feed-forward neural network.
    [Show full text]