DOCUMENT RESUME FL 024 768 ED 413 747 de Sopena, Luis AUTHOR Speech Recognition: A General Overview. TITLE 1995-00-00 PUB DATE 6p.; In: Language Resources for Language Technology: NOTE Proceedings of the TELRI (Trans-European Language Resources Infrastructure) European Seminar (1st, Tihany, Hungary, September 15-16, 1995); see FL 024 759. Speeches/Meeting Papers (150) Reports Descriptive (141) PUB TYPE MF01/PC01 Plus Postage. EDRS PRICE *Computational Linguistics; *Computer Software; *Discourse DESCRIPTORS Analysis; Foreign Countries; Information Technology; *Language Processing; *Language Research; Linguistic Theory; *Oral Language; Trend Analysis *Speech Recognition IDENTIFIERS ABSTRACT Speech recognition is one of five main areas in the field of speech processing. Difficulties in speech recognition include variability in sound within and across speakers, in channel, in background noise, and of speech production. Speech recognition can be used in a variety of situations: to perform query operations and phone call transfers; for data entry; for command and control operations; and in dictation. Technical characteristics of speech recognition systems depend on several variables, the most important of which are vocabulary size, speaker dependence, speaker mode, domain dependence, and multiple language support. Knowledge sources are based on three models: set of phonemes (acoustic); word lexicon; and language. The objective of the speech recognition process is to determine the sequence of words that most probably caused the observed sequence of acoustic vectors. Currently, speech recognition systems can recognize a large number of words, recognize discrete speech, handle 70-100 words per minute, and handle several languages with a high recognition rate. In the future, speech recognition systems will be able to handle any speaker without need for training, continuous speech, very large vocabularies, telephone communication, and natural language understanding. (MSE) ******************************************************************************** Reproductions supplied by EDRS are the best that can be made from the original document. ******************************************************************************** Speech Recognition A General Overview Luis de Sopetia IBM S. A. Madrid Scientific Centre Santa Hortensia 26-28 E-280 002 Madrid Tel.: +341 397 5752 Fax: +341 519 3990 E-mail: [email protected] PERMISSION TO REPRODUCE AND OF EDUCATION U.S DEPARTMENT Research and Improvement DISSEMINATE THIS MATERIAL Off ice of Educational RESOURCES INFORMATION EDUCATIONAL HAS BEEN GRANTED BY CENTE as has been reproduced IQ This document or organization received from the person originating it. been made to Minor changes have quality. improve reproduction this TO THE EDUCATIONAL RESOURCES opinions stated in Points of view or necessarily represent INFORMATION CENTER (ERIC) document do not policy. official OERI position or . C BEST COPY AVAILABLE Luis de Sope iia 100 1. Areas of Speech Processing There are five main areas in the field of Speech Processing: 1) Speech Coding deals with the compression of the digital representation of the speech signal in order to facilitate economical transmission or storage. 2) In Speech Synthesis, a synthetic speech signal is created from preexisting text with an attempt at reaching maximum intelligibility and naturalness. 3) Using techniques for Speaker Identification, the machine identifies the speaker by his/her voice in order to ensure restricted access to information, computer, or the physical premises. 4) In Speech Recognition, the information in a spoken message is identified so as to have the computer perform'the corresponding command or transcribe in written form the dictated text. 5) Finally, Spoken Language Translation deals with two-way communication via speech: a spoken message is identified, translated into a different language and this translation synthesised in speech form, in order, e.g., to enable a dialogue between speakers of different languages. 2. Difficulties in Speech Recognition There are some well-known difficulties in the field of speech recognition, shown in the list below: The variability of sounds (words, phrases, subword units), within a single speaker and across different speakers. The variability of channel, depending on the characteristics of the different types of microphones. The variability of background noise: side conversations, street noise, telephone rings, etc. The variability of speech production, which adds spurious sounds to words proper (mouth clicks, hesitations, breath noise.) 3. Main Functions of Speech Recognition Speech recognition can be used in a variety of situations: 1) To perform Query operations, such as the consultation via telephone of a bank for account balances, the consultation of phone information lines for theatre schedules and the like, and also for phone call transfers. Speech Recognition 101 2) Data entry situations may include the giving of a credit card number, dialing from mobile phones, and booking airline out filling forms, reservations. 3) Command and Control operations in which speech recognition is im- portant occur when the hands and/or eyes are busy, during menu navigation and machine control, and while completing dark room work. 4) Speech- recognition plays a key role in dictation when entering free text into a computer via speech. 4. Technical Characteristics of Speech Recognition Systems The technical characteristics of speech recognition systems depend on several variables, the most important of which are the following: 1) The vocabulary size can range from small (10-100 words) for simple commands, to medium (1000 words) for form filling, or to large (more than 20 000) for such complex situations as dictation. 2) Other than vocabulary size, the speaker dependence of a given system can vary from being trained to a specific speaker, to being adaptive to each user as (s)he speaks, or even speaker independent. 3) The speaking mode varies between continuous text and isolated words, where pauses between words are needed for an adequate recognition. 4) Speech recognition systems can be domain dependent, meaning they can only recognize a constrained syntax (e.g., a list of commands or of questions), or independent, where free text can be dictated. 5) Multiple language support is also an important characteristic. 5. Knowledge Sources in Speech Recognition The knowledge sources in speech recognition are based on three different models: 1) Set of Phoneme Models: Reference to the typical sound of a phoneme, specified by the probability distribution of its spectral and temporal properties. 2) Word Lexicon: Represented as a sequence of the above phonemes (Acoustic Model). 3) Language Model: Statistical model extracted from large corpora of texts. Luis de Sopena 102 6. Speech Recognition Process The objective of the speech recognition process is to determine the sequence of words which caused most probably the observed sequence of acoustic vectors (see figure 1). Speech Recognition Process Speech Input Feature Extraction Spectral Representation Phoneme Acoustic-Phonetic Models Processor Phonetic Units Lexicon Word Search Word Lattice Language "Syntax" Model Analysis Recognised Text Figure 1. Determine sequence of words which caused MOST PROBABLY the observed sequence of acoustic vectors 5 Speech Recognition 103 7. Speech Recognition Today and Tomorrow An example of a present-day Speech Recognition system is the IBM Voicepe Dictation System. Its most important characteristics are: Works on a 486 SX 25 Recognises more than 30K different words Needs a short enrollment process Recognises discrete speech (with small pauses between words) Able to handle 70-100 words per minute - Available for 6 languages With a very high recognition rate (>96%) Tomorrow, however, research is promising much more. Speech recognition systems will be able to handle: Any speaker, without need for training Continuous speech Very large vocabularies (more than 250K words) With telephone capabilities - Including natural language understanding On Personal Digital Assistants These systems will be used in dictation, phone mail, DB access, home shopping, translation, and much more. Most important of all, Speech will be an "enabler", i.e., existing and new applications will be accessible using speech. 8. Main Players in the Field of Speech Recognition The main players in the field of speech recognition are the following: the European Community ARPA (Wall Street Journal Contest, Air Travel Info Service (ATIS)) Industrial Research (IBM in dictation, AT & T for phone services, and many smaller companies) One of the continual points of discussion in the field of speech recognition is the relative importance of English as compared to other languages. But nonetheless, speech systems are developed for other major languages as well (e.g., French, German, Spanish). E7 a) U.S. Department of Education Office of Educational Research and Improvement (OERI) Educational Resources Information Center (ERIC) REPRODUCTION RELEASE (Specific Document) I. DOCUMENT IDENTIFICATION: Title: TELRI - Proceedings of the First European Seminar7"Language Resources for:Lan- guage Technology", Tihany, Hungary, Sept. 15 and 16, 1995 Author(s): Heike Rettig ..(Ed .) Corporate Source: Publication Date: 1996 II. REPRODUCTION RELEASE: In order to disseminate as widely as possible timely and significant materials of interest to the educational community, documents announced in the monthly abstract journal of the ERIC system, Resources in Education (RIE), are usually made available to users in microfiche, reproduced paper copy, and electronic/optical media, and sold through the ERIC Document Reproduction Service (EDRS) or other ERIC vendors. Credit is given to the source of each document, and, if reproduction release is granted, one of the following notices is affixed to the document. If permission is granted to reproduce and disseminate the identified document, please CHECK ONE of the following two options and sign at the bottom of the page. The sample sticker shown below will be The sample sticker shown below will be affixed to all Level 2 documents affixed to all Level 1 documents LI PERMISSION TO REPRODUCE AND PERMISSION TO REPRODUCE AND I DISSEMINATE THIS DISSEMINATE THIS MATERIAL MATERIAL IN OTHER THAN PAPER HAS BEEN GRANTED BY COPY HAS BEEN GRANTED BY Check here Check here For Level 1 Release: For Level 2 Release: Permitting reproduction in Permitting reproduction in microfiche (4' x 6" film) or microfiche (4" x 6' film) or TO THE EDUCATIONAL RESOURCES other ERIC archival media other ERIC archival media TO THE EDUCATIONAL RESOURCES INFORMATION CENTER (ERIC) (e.g., electronic or optical) INFORMATION CENTER (ERIC) (e.g., electronic or optical), and paper copy. but not in paper copy. Level 1 Level 2 Documents will be processed as indicated provided reproduction quality permits. If permission to reproduce is granted, but neither box is checked, documents will be processed at Level 1. 'I hereby grant to the Educational Resources Information Center (ERIC) nonexclusive permission to reproduce and disseminate this document as indicated above. Reproduction from the ERIC microfiche or electronic/optical media by persons other than ERIC employees and its system contractors requires permission from the copyright holder. Exception is made for non-profit reproduction by libraries and other service agencies to satisfy information needs of educators in response to discrete inquiries. " Sign Signature: Printed Name/Position/Title: here-) Norbert Volz, M.A. please Organization/Address: farignel'e:feGt-PleflegTAX: deutsche So ache Institut COI. +49...E21 68161 Mannheim E-Mail Address: Date: P 5, 6-1 3 68016 Mannheim volz(at)ids-mannhei0.de Pc:Afach 101621 28/11/97 (over)