EP0919052B1

EP0919052B1 - A method and a system for speech-to-speech conversion

Info

Publication number: EP0919052B1
Application number: EP97919840A
Authority: EP
Inventors: Bertil Lyberg
Original assignee: Telia AB
Current assignee: Telia AB
Priority date: 1996-05-13
Filing date: 1997-04-08
Publication date: 2003-07-09
Anticipated expiration: 2017-04-08
Also published as: DE69723449D1; DK0919052T3; DE69723449T2; NO318557B1; SE506003C2; EP0919052A1; SE9601811D0; NO985179L; WO1997043756A1; NO985179D0; SE9601811L

Description

The invention relates to a speech-to-speech conversion system and method which are capable of matching the dialect of speech outputs to that of the respective speech inputs, and to a voice responsive communication system including a speech-to-speech conversion system and operating in accordance with a speech-to-speech conversion method.

In known voice responsive communication systems, the speech information, which is stored in a database and used to provide appropriate synthesised spoken responses to voice inputs utilising a speech-to-speech conversion system, is normally reproduced in a dialect which conforms to a standard national dialect. Thus, when there are significant differences between the dialect of the voice inputs and the standard national dialect, it may prove difficult, in certain circumstances, for the database of known voice responsive communication systems to interpret received speech information, i.e. the voice inputs. It may also be difficult for the person making the voice inputs to fully understand the spoken response. Even if such responses are understandable to a recipient, it would be more user friendly if the dialect of the spoken response is the same as the dialect of the related voice input.

Also, with artificial reproduction of a spoken language, there is a need for the language to be reproduced naturally and with the correct accentuation. In particular, the meaning of a word can have widely different meanings depending on language stress. Also, the meaning of one and the same sentence can be given a different significance depending on where the stress is placed. Furthermore, the stressing of sentences, or parts thereof, determines sections which are emphasised in the language and which may be of importance in determining the precise meaning of the spoken language.

The need for artificially produced speech to be as natural as possible and have the correct accentuation is of particular importance in voice responsive communication devices and/or systems which produce speech in various contexts. With known voice responsive arrangements, the reproduced speech is difficult to understand and interpret. There is, therefore, a need for a speech-to-speech conversion system in which the artificial speech outputs are natural, have the correct accentuation, and are readily understandable.

With languages having well developed sentence accent stress and/or pitch in individual words, identification of the natural meaning of the words/sentences is very difficult. The fact that stresses can be incorrectly placed increases the risk of misinterpretation, or that the meaning is completely lost for the listening party.

Various types of speech recognition systems are known. It is usual, in such systems, for the speech recognition equipment to be trained to recognise speech from a large number of persons. Also, the speech training follows a particular dialect, or dialects. The information collected through this process is then used by the system to interpret incoming speech. Thus, such systems cannot normally recognise dialect variations in speech which are outside the particular dialect(s) for which the system is trained.

As an example, document WO-A-96/00962 discloses a speech recognition system for recognizing dialectal variations in a language.

In languages with tone word accents and tone language, the intonation pattern of the language constitutes a very important part in the understanding of the language, but known systems take no account of these speech characteristics. As a consequence of this, the recognition of words and phrases, with known speech recognition systems, can give rise to misinterpretations. The known speech-recognition systems which are adapted to take account of dialects in speech, are specifically tailored for a particular dialect and are not, therefore, adapted to recognise different dialects in a language.

In future, speech recognition equipments will, to an ever increasing extent, be used in many different applications where there is a need to be able to recognise different dialects in a language. The dialectal variations in a language have, in the past, been difficult to determine and, as a consequence of this, difficulties have been experienced in obtaining a proper understanding of artificially produced speech. Furthermore, the known speech recognition equipments cannot generally be used with different languages.

Thus, whilst known speech recognition systems are adapted to recognise, through training, a particular dialect in a language, it is not possible, for such systems, to recognise different dialects in that language, or dialects in different languages, using the same speech recognition equipment, without further training.

The artificial interpretation of speech has, therefore, been difficult, or impossible, to perform with known speech recognition equipments, due to the inability of such systems to recognise dialectal variations.

Apart from the technical problems of correctly interpreting speech, it is necessary, in voice responsive/control systems, for the verbal instructions, or commands, to be correctly interpreted, otherwise it would not be possible to provide proper responses, or effect proper control of different types of equipments, and/or services, for example, in a telecommunication network.

In order to overcome these difficulties, it would be necessary for a voice responsive communication system to be capable of interpreting the received speech information, irrespective of dialect, and to match the dialect of speech outputs to that of the respective speech inputs. Also, in order to be able to determine the meaning of single words, or phrases, in an unambiguous manner in a spoken sequence, it would be necessary for the speech-to-speech converters used in a voice responsive communication system to be capable of determining, and taking account of, stresses in the spoken sequence.

It is an object of the present invention to provide a system and method for speech-to-speech conversion which are capable of matching the dialect of speech outputs to that of the respective speech inputs.

It is another object of the present invention to provide a system and method for speech-to-speech conversion which are adapted to recognise and interpret speech inputs and, in particular, the dialect, sentence accent and stressing of spoken sequences, using the fundamental tone curve of the spoken sequences.

It is a further object of the present invention to provide a voice responsive communication system including a speech-to-speech conversion system which is capable of matching the dialect of speech outputs to that of the respective speech inputs.

The invention as claimed in claims 1-26 provides a speech-to-speech conversion system for providing, at the output thereof, spoken responses to speech inputs to the system including speech recognition means for the input speech; interpretation means for interpreting the content of the recognised input speech; and a database containing speech information data for use in the formulation of said spoken responses, the output of said interpretation means being used to access said database and obtain speech information data therefrom, characterised in that the system further includes extraction means for extracting prosody information from the input speech; means for obtaining dialectal information from said prosody information; and text-to-speech conversion means for converting the speech information data obtained from said database into a spoken response using said dialectal information, the dialect of the spoken response being matched to that of the input speech.

The speech recognition means may be adapted to identifying a number of phonemes from a segment of the input speech and to interpret the phonemes, as possible words, or word combinations, to establish a model of the speech, the speech model having word and sentence accents according to a standardised pattern for the language of the input speech.

The prosody information extracted from the input speech is preferably the fundamental tone curve of the input speech. In this case, the means for obtaining dialectal information from said prosody information includes first analysing means for determining the intonation pattern of the fundamental tone of the input speech and thereby the maximum and minimum values of the fundamental tone curve and their respective positions; second analysing means for determining the intonation pattern of the fundamental tone curve of the speech model and thereby the maximum and minimum values of the fundamental tone curve and their respective positions; comparison means for comparing the intonation pattern of the input speech with the intonation pattern of the speech model to identify a time difference between the occurrence of the maximum and minimum values of the fundamental tone curves of the incoming speech in relation to the maximum and minimum values of the fundamental tone curve of the speech model, the identified time difference being indicative of dialectal characteristics of the input speech. The time difference may be determined in relation to an intonation pattern reference point, for example, the point at which a consonant/vowel limit occurs.

The speech-to-speech conversion system may include means for obtaining information on sentence accents from said prosody information. In which case, the speech recognition means includes checking means for lexically checking the words in the speech model and for syntactically checking the phrases in the speech model, the words and phrases which are not linguistically possible being excluded from the speech model. The checking means are, with this arrangement, adapted to check the orthography and phonetic transcription of the words in the speech model, the transcription information including lexically abstracted accent information, of type stressed syllables, and information relating to the location of secondary accent. The accent information may, for example, relate to tonal word accent I and accent II.

The sentence accent information and/or sentence stressing may be used, to advantage, in the interpretation of the content of the recognised input speech.

The speech-to-speech conversion system may include dialogue management means for managing a dialogue with the database, said dialogue being initiated by the interpretation means. In a preferred arrangement, the dialogue with the database results in the application of speech information data to the text-to-speech conversion means.

The invention also provides, in a voice responsive communication system, a method for providing a spoken response to a speech input to the system, said response having a dialect to match that of the speech input, said method including the steps of recognising and interpreting the input speech, and utilising the interpretation to obtain speech information data from a database for use in the formulation of said spoken response, characterised in that said method further includes the steps of extracting prosody information from the input speech, obtaining dialectal information from said prosody information, and converting the speech information data obtained from said database into said spoken response using said dialectal information.

The recognition and interpretation of the input speech includes the steps of identifying a number of phonemes from a segment of the input speech and interpreting the phonemes, as possible words, or word combinations, to establish a model of the speech, the speech model having word and sentence accents according to a standardised pattern for the language of the input speech.

In a preferred method, the prosody information extracted from the input speech is the fundamental tone curve of the input speech. In this case, the method according to the present invention includes the steps of determining the intonation pattern of the fundamental tone of the input speech and thereby the maximum and minimum values of the fundamental tone curve and their respective positions; determining the intonation pattern of the fundamental tone curve of a speech model and thereby the maximum and minimum values of the fundamental tone curve and their respective positions; comparing the intonation pattern of the input speech with the intonation pattern of the speech model to identify a time difference between the occurrence of the maximum and minimum values of the fundamental tone curves of the incoming speech in relation to the maximum and minimum values of the fundamental tone curve of the speech model, the identified time difference being indicative of dialectal characteristics of the input speech. The time difference may be determined in relation to an intonation pattern reference point, for example, the point at which a consonant/vowel limit occurs.

The method, according to the present invention, may include the step of obtaining information on sentence accents from said prosody information. In accordance with this method, the words in the speech model are checked lexically and the phrases in the speech model are checked syntactically, the words and phrases which are not linguistically possible being excluded from the speech model. Also, in accordance with this method, the orthography and phonetic transcription of the words in the speech model may be checked, the transcription information including lexically abstracted accent information, of type stressed syllables, and information relating to the location of secondary accent. The accent information may relate to tonal word accent I and accent II.

In accordance with the method of the present invention, sentence accent information and/or sentence stressing, may be used in the interpretation of the content of the recognised input speech.

The method according to the present invention may include the step of initiating a dialogue with the database to obtain speech information data for formulating said spoken response, said dialogue being initiated following the interpretation of the input speech. The dialogue with the database may result in the application of speech information data to text-to-speech conversion means.

The invention further provides a voice responsive communication system which includes a speech-to-speech conversion system as outlined in the preceding paragraphs, or utilises a method as outlined in the preceding paragraphs for providing a spoken response to a speech input to the system.

In essence, the characteristic features of the speech-to-speech conversion system and method, according to the present invention, are that:

prosody information is extracted from speech, applied to the input of the system, and handled by the method;
the prosody information is in the form of the fundamental tone curve of the input speech;
the fundamental tone curve is used to obtain dialectal, sentence accent and sentence, stressing information for the input speech;
the sentence accent and stressing information is used in the interpretation of the speech inputs, the result of the interpretation being used to obtain speech information data from a database which is used in the formulation of voice responses to the speech inputs; and
the dialectal information is used to ensure that the voice responses to the speech inputs have a dialect to match that of respective speech inputs.

The foregoing and other features according to the present invention will be better understood from the following description with reference to the single figure of the accompanying drawings which illustrates, in the form of a block diagram, a speech-to-speech conversion system according to the present invention.

It will be seen from the single figure of the accompanying drawings that a speech-to-speech conversion system, according to the present invention, includes, at the input 1 thereof, a speech recognition unit 2 and an extraction unit 3 for extracting prosody information from speech applied to the system input 1, i.e. the fundamental tone curve of the input speech. Thus, speech inputs, applied to the input 1, are simultaneously applied to the

units

2 and 3.

The output of the speech recognition unit 2 and an output of the extraction unit 3 are connected to separate inputs of an interpretation unit 4, the output of which is connected to a database management unit 5. The database management unit 5 which is adapted for two way communication with a database 6, is connected at the output thereof to the input of a text-to-speech converter 7. The dialogue between the database 6 and the database management unit 5 can be effected by any known database communication language, for example, SQL (Structured Query Language). The output of the text-to-speech converter 7 provides a synthesised speech output for the speech-to-speech conversion system.

As shown in the single figure of the drawings, a further output of the extraction unit 3 is connected to the input of a prosody analyzer unit 8 which is adapted for two way communication with the text-to-speech converter 7. The prosody analyzer unit 8 is adapted, as a part of the text-to-speech conversion process of the converter 7, to analyze the prosody information, i.e. the fundamental tone curve, of the synthesised speech and make any necessary corrections to the intonation pattern of the synthesised speech in accordance with the dialectal information extracted from the input speech. Thus, the dialect of the synthesised speech output of the speech-to-speech conversion system will match that of the input speech.

It will, therefore, be seen from the foregoing that the present invention is adapted to provide a spoken response to a speech input to the speech-to-speech conversion system which has a dialect to match that of the speech input and that this conversion process includes the steps of recognising and interpreting the input speech, utilising the interpretation to obtain speech information data from a database for use in the formulation of the spoken response, extracting prosody information from the input speech, obtaining dialectal information from the prosody information, and converting the speech information data obtained from said database into the spoken response using the dialectal information. The manner in which this can be effected will be outlined in the following paragraphs.

In practice, the speech inputs to the speech-to-speech conversion system, which may be in many forms, for example, requests for information on particular topics, such as banking or telephone services, or general enquiries concerning such services, are applied to the input 1 and thereby to the inputs of the

units

2 and 3.

The speech recognition unit 2 and interpretation unit 4 are adapted to operate, in a manner well known to persons skilled in the art, to recognise and interpret the speech inputs to the system. The speech recognition unit 2 may, for example, operate by using a Hidden Markov model, or an equivalent speech model. In essence, the function of the

units

2 and 4 is to convert speech inputs to the system into a form which is a faithful representation of the content of the speech inputs and suitable for application to the input of the database management unit 5. In other words, the content of the textual information data at the output of the interpretation unit 4 must be an accurate representation of the speech input and be usable by the database management unit 5 to access, and extract speech information data from, the database 6 for use in the formulation of a synthesised spoken response to the speech input. In practice, this process would, in essence, be effected by identifying a number of phonemes from a segment of the input speech which are combined into allophone strings, the phonemes being interpreted as possible words, or word combinations, to establish a model of the speech. The established speech model will have word and sentence accents according to a standardised pattern for the language of the input speech.

The information, concerning the recognised words and word combinations, generated by the speech recognition unit 2, may, in practice, be checked both lexically (using a lexicon, with orthography and transcription) and syntactically. The purpose of these checks is to identify and exclude any words which do not exist in the language concerned, and/or any phrase whose syntax does not correspond with the language concerned.

Thus, in accordance with the present invention, the speech recognition unit 2 ensures that only those words, and word combinations, which are found to be acceptable both lexically and syntactically, are used to create a model of the input speech. In practice, the intonation pattern of the speech model is a standardised intonation pattern for the language concerned, or an intonation pattern which has been established by training, or explicit knowledge, using a number of dialects of the language concerned.

The prosody information, i.e. the fundamental tone curve, extracted from the input speech by the extraction unit 3, can be used to obtain dialectal, sentence accent and sentence stressing, information, for use by the speech-to-speech conversion system and method of the present invention. In particular, the dialectal information can be used by the speech-to-speech conversion system and method to match the dialect of the output speech to that of the input speech and the sentence accent and stressing information can be used in the recognition and interpretation of the input speech.

In accordance with the present invention, the means for obtaining dialectal information from the prosody information includes;

- first analysing means for determining the intonation pattern of the fundamental tone of the input speech and thereby the maximum and minimum values of the fundamental tone curve and their respective positions;

second analysing means for determining the intonation pattern of the fundamental tone curve of the speech model and thereby the maximum and minimum values of the fundamental tone curve and their respective positions; and
comparison means for comparing the intonation pattern of the input speech with the intonation pattern of the speech model to identify a time difference between the occurrence of the maximum and minimum values of the fundamental tone curves of the incoming speech in relation to the maximum and minimum values of the fundamental tone curve of the speech model, the identified time difference being indicative of the dialectal characteristics of the input speech.

The time difference, referred to above, may be determined in relation to an intonation pattern reference point.

In the Swedish language, the-difference, in terms of intonation pattern, between different dialects can be described by different points in time for word and sentence accent, i.e. the time difference can be determined in relation to an intonation pattern reference point, for example, the point at which a consonant/vowel limit occurs.

Thus, in a preferred arrangement for the present invention, the reference against which the time difference is measured, is the point at which the consonant/vowel boundary, i.e. the CV-boundary, occurs.

The identified time difference which, as stated above, is indicative of the dialect in the input speech, i.e. the spoken language, is applied to the text-to-speech converter 7 to enable the intonation pattern, and thereby the dialect, of the speech output of the system to be corrected so that it corresponds to the intonation pattern of the corresponding words and/or phrase of the input speech. Thus, this corrective process enables the dialectal information in the input speech to be incorporated into the output speech.

As stated above, the fundamental tone curve of the speech model is based on information resulting from the lexical (orthography and transcription) and syntactic checks. In addition, the transcription information includes lexically abstracted accent information, of type stressed syllables, i.e. tonal word accents I and II, and information relating to the location of secondary accent, i.e. information given, for instance, in dictionaries. This information can be used to adjust the recognition pattern of the speech recognition model, for example, the Hidden Markov model, to take account of the transcription information. A more exact model of the input speech is, therefore, obtained during the interpretation process.

A further consequence of this- speech model corrective process is that, through time, the speech model will have an intonation pattern which has been established by a training process.

Also, with the system and method of the present invention, the speech model is compared with a spoken input sequence, and any difference there between can be determined and used to bring the speech model into conformity with the spoken sequence and/or to determine stresses in the spoken sequence.

The identification of the stresses in a spoken sequence makes it possible to determine the precise meaning of the spoken sequence in an unambiguous manner. In particular, relative sentence stresses can be determined by classifying the ratio between variations and declination of the fundamental tone curve, whereby emphasised sections, or individual words can be determined. In addition, the pitch of the speech can be determined from the declination of the fundamental tone curve.

Thus, in order to take account of sentence stresses in the recognition and interpretation of the speech inputs to the speech-to-speech conversion system of the present invention, the extraction unit 3, in association with the interpretation unit 4, is adapted to determine:

a first ratio between the variation and declination of the fundamental tone curve of the input speech;
a second ratio between the variation and declination of the fundamental tone curve of the speech model; and
comparing the first and second ratios, any identified difference being used to determine sentence accent placements.

In addition, classification of the ratio between the variation and declination of the fundamental tone curve, makes it possible to identify/determine relative sentence stresses, and emphasised sections, or words.

Also, the relation between the variation and declination of the fundamental tone curve can be utilised to determine the dynamic range of the fundamental tone curve.

The information obtained in respect of the fundamental tone curve concerning dialect, sentence accent and stressing, can be used for the interpretation of speech by the interpretation unit 4, i.e. the information can be used, in the manner outlined above, to obtain a better understanding of the content of the input speech and bring the intonation pattern of the speech model into conformity with the input speech.

Since the corrected speech model exhibits the language characteristics (including dialect information, sentence accent and stressing) of the input speech it can be used to give an increased understanding of the input speech and be effectively used by the database management unit 5 to obtain the required speech information data from the database 6 to formulate a response to a voice input to the speech-to-speech conversion system.

The ability to readily interpret different dialects in a language using fundamental tone curve information, is of some significance because such interpretations can be effected without having to train the speech recognition system. The result of this is that the size, and thereby cost, of a speech recognition system, made in accordance with the present invention, can be much smaller than would be possible with known systems. These are, therefore, distinct advantages over known speech recognition systems.

Also, the ability to detect speech, irrespective of dialect variations, in accordance with the system and method of the present invention, makes it possible to use speech in many different voice-responsive applications.

The system is, therefore, adapted to recognise and accurately interpret the content of speech inputs and to tailor the dialect of the voice response to match the dialect of the voice input. This process provides a user friendly system because the language of the man-machine dialogue is in accordance with the dialect of the user concerned.

The present invention is not limited to the embodiments outlined above, but can be modified within the scope of the appended patent claims.

Claims

A speech-to-speech conversion system for providing, at the output thereof, spoken responses to speech inputs to the system including speech recognition means for the input speech; interpretation means for interpreting the content of the recognised input speech; and a database containing speech information data for use in the formulation of said spoken responses, the output of said interpretation means being used to access said database and obtain speech information data therefrom, characterised in that the system further includes extraction means for extracting prosody information from the input speech; means for obtaining dialectal information from said prosody information; and text-to-speech conversion means for converting the speech information data obtained from said database into a spoken response using said dialectal information, the dialect of the spoken response being matched to that of the input speech, the means for obtaining dialectal information from said prosody information includes first analysing means for determining the intonation pattern of the fundamental tone of the input speech and thereby the maximum and minimum values of the fundamental tone curve and their respective positions; second analysing means for determining the intonation pattern of the fundamental tone curve of the speech model and thereby the maximum and minimum values of the fundamental tone curve and their respective positions; comparison means for comparing the intonation pattern of the input speech with the intonation pattern of the speech model to identify a time difference between the occurrence of the maximum and minimum values of the fundamental tone curves of the incoming speech in relation to the maximum and minimum values of the fundamental tone curve of the speech model, the identified time difference being indicative of dialectal characteristics of the input speech.
A speech-to-speech conversion system as claimed in claim 1, characterised in that the speech recognition means are adapted to identifying a number of phonemes from a segment of the input speech and include interpretation means for interpreting the phonemes, as possible words, or word combinations, to establish a model of the speech, the speech model having word and sentence accents according to a standardised pattern for the language of the input speech.
A speech-to-speech conversion- system as claimed in claim 2, characterised in that the prosody information extracted from the input speech is the fundamental tone curve of the input speech.
A speech-to-speech conversion system as claimed in claim 3, characterised in that the time difference is determined in relation to an intonation pattern reference point.
A speech-to-speech conversion system as claimed in claim 4 characterised in that intonation pattern reference point, against which the time difference is measured, is the point at which a consonant/vowel limit occurs.
A speech-to-speech conversion system as claimed in any one of the preceding claims, characterised in that the system further includes means for obtaining information on sentence accents from said prosody information.
A speech-to-speech conversion system as claimed in claim 6, characterised in that the speech recognition means includes checking means for lexically checking the words in the speech model and for syntactically checking the phrases in the speech model, the words and phrases which are not linguistically possible being excluded from the speech model, in that the checking means are adapted to check the orthography and phonetic transcription of the words in the speech model, in that the transcription information includes lexically abstracted accent information, of type stressed syllables, and information relating to the location of secondary accent.
A speech-to-speech conversion system as claimed in claim 7, characterised in that the accent information relates to tonal word accent I and accent II.
A speech-to-speech conversion system as claimed in any one of claims 6 to 8, characterised in that said sentence accent information is used in the interpretation of the content of the recognised input speech.
A speech-to-speech conversion system as claimed in any one of the preceding claims, characterised in that sentence stresses are determined and used in the interpretation of the content of the recognised input speech.
A speech-to-speech conversion system as claimed in any one of the preceding claims, characterised in that the system further includes dialogue management means for managing a dialogue with the database, said dialogue being initiated by the interpretation means.
A speech-to-speech conversion system as claimed in claim 11, characterised in that the dialogue with the database results in the application of speech information data to the text-to-speech conversion means.
A speech-to-speech conversion system as claimed in claim 10, or claim 11, characterised in that the dialogue with the database is effected using SQL.
A voice responsive communication system including a speech-to-speech conversion system as claimed in any one of the preceding claims.
A method for providing a spoken response to a speech input to a voice responsive communication system, said response having a dialect to match that of the speech input, said method including the steps of recognising and interpreting the input speech, and utilising the interpretation to obtain speech information data from a database for use in the formulation of said spoken response, characterised in that said method further includes the steps of extracting prosody information from the input speech, obtaining dialectal information from said prosody information, and converting the speech information data obtained from said database into said spoken response using said dialectal information, the steps of determining the intonation pattern of the fundamental tone of the input speech and thereby the maximum and minimum values of the fundamental tone curve and their respective positions; determining the intonation pattern of the fundamental tone curve of a speech model and thereby the maximum and minimum values of the fundamental tone curve and their respective positions; comparing the intonation pattern of the input speech with the intonation pattern of the speech model to identify a time difference between the occurrence of the maximum and minimum value of the fundamental tone curves of the incoming speech in relation to the maximum and minimum values of the fundamental tone curve of the speech model, the identified time difference being indicative of dialectal characteristics of the input speech.
A method as claimed in claim 15, characterised in that the recognition and interpretation of the includes the steps of identifying a number of phonemes from a segment of the input speech and interpreting the phonemes, as possible words, or word combinations, to establish a model of the speech, the speech model having word and sentence accents according to a standardised pattern for the language of the input speech.
A method as claimed in claim 16, characterised in that the prosody information extracted from the input speech is the fundamental tone curve of the input speech.
A method as claimed in claim 15, characterised in that the time difference is determined in relation to an intonation pattern reference point.
A method as claimed in claim 18, characterised in that the intonation pattern reference point, against which the time difference is measured, is the point at which a consonant/vowel limit occurs.
A method as claimed in any one of the claims 15 to 19, characterised by the step of obtaining information on sentence accents from said prosody information.
A method as claimed in claim 20, characterised in that the words in the speech model are checked lexically, in that the phrases in the speech model are checked syntactically, in that the words and phrases which are not linguistically possible are excluded from the speech model, in that the orthography and phonetic transcription of the words in the speech model are checked, and in that the transcription information includes lexically abstracted accent information, of type stressed syllables, and information relating to the location of secondary accent.
A method as claimed in claim 21, characterised in that the accent information relates to tonal word accent I and accent II.
A method as claimed in any one of claims 20 to 22, characterised by the step of using said sentence accent information in the interpretation of the input speech.
A method as claimed in any one of the claims 15 to 23, characterised by the step of initiating a dialogue with the database to obtain speech information data for formulating said spoken response, said dialogue being initiated following the interpretation of the input speech.
A method as claimed in claim 24, characterised in that the dialogue with the database results in the application of speech information data to text-to-speech conversion means.
A voice responsive communication system which is adapted to utilise a method as claimed in any one of the claims 15 to 25 for providing a spoken response to a speech input to the system.