Classifications

• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L13/00—Speech synthesis; Text to speech systems
  • G10L13/08—Text analysis or generation of parameters for speech synthesis out of text, e.g. grapheme to phoneme translation, prosody generation or stress or intonation determination
  • G10L13/10—Prosody rules derived from text; Stress or intonation

Definitions

• the present invention refers to a method of implementing intonation curves for vocal messages, and to a speech synthesis method and system using the same.
• vocal messages are very common in many technical applications, for example for telephone services, such as the "right time of day” and the "news”, and at the airports and stations. Till some years ago such vocal messages were pronounced by physical persons; afterwards it became popular to pre-record whole vocal messages, for example on magnetic media, and to emit them by selection.
• Today tendency is to use automatic speech synthesis systems which, using a vocabulary of pre-recorded words (or syllables or groups of letters), synthesize the vocal message concatenating such elementary entities according to the correspondent text messages.
• the quality of such automatic speech synthesis systems is evaluated on the base of the naturalness of the emitted vocal message as can be appreciated by a listener; the naturalness depends on, in addition to other elements, the emission intonation and the length of the single words and, as a consequence, of the whole vocal message.
• the methods of the present invention are based on the idea of handling independently and separately the pronunciation and the intonation of vocal messages.
• the methods of the present invention make use of a small database including word intonation curves of words, extract from such database the appropriate (from the intonation point of view) word intonation curves corresponding to the words of the vocal message to be synthesized, and concatenate them to form the message intonation curve.
• the word intonation database includes word intonation curves only of a limited number of words, and the methods extract from such database the word intonation curves corresponding, only from the intonation point of view, to the words of the vocal message to be synthesized.
• An important advantage of the present invention is that it is applicable to many different languages with minor changes which are very simple and straightforward for the person skilled in the art of speech processing.
• the methods of the present invention may be used only if the vocal messages are coded through a technique that permits the instantaneous control of intonation, i.e. of pitch, for example the LPC technique.
• the majority of these techniques divides the vocal signal into a sequence of frames and codifies each frame through a set of parameters, one of them being or being directly related to the pitch.
• the methods of implementing a message intonation curve for a vocal message which will be described, are embodiments of the idea of generating the intonation independently and separately from the generation of the pronunciation of the vocal message.
• a first method, according to the present invention, of implementing a message intonation curve for a vocal message corresponding to a text message consisting of the concatenation of words makes use of a word intonation database including word intonation curves and includes the steps of:
• the message intonation curve is the concatenation of the word intonation curves so obtained.
• the word intonation curves could be sequences of values corresponding to instantaneous intonation or pitch.
• the message intonation curve depends, in addition, on the general intonation of the text message, i.e. interrogative, affirmative, exclamatory, dubitative.
• step a) the value of a second parameter is determined corresponding to the general intonation of the text message, and in step b) the word intonation curve corresponds to the word and to the values of two parameters.
• a number may be determined which represents the maximum number of syllables of a word in that particular language; naturally a lower number may be chosen (predetermined limit) and as a consequence only a certain percentage of the words of that particular language fall within such limit. In an embodiment of the present invention for the Italian language it turned out that choosing this predetermined limit equal to six will lead to satisfactory results.
• the words of a language may be divided in a number of classes each corresponding to a different number of syllables. Each class may be further divided according to the position of the accented syllable (primary accent).
• the number of classes so obtained is twenty-two : two classes for the words composed of one syllable, two classes for two syllables, three classes for three syllables, ..., six classes for six syllables.
• the fact that two classes are necessary for one syllable depends from the fact that the Italian words "il” and "la" (one-syllable words) have two different intonations as the vowel in the first case is followed by a consonant and not in the second case.
• its correct intonation in a sentence is related to its position inside it; at least three positions should be considered corresponding to: the beginning, the middle, and the end of the sentence.
• a second method, according to the present invention, of implementing a message intonation curve for a vocal message corresponding to a text message consisting of the concatenation of words makes use of a word intonation database including word intonation curves and includes the steps of:
• the message intonation curve is the concatenation of the word intonation curves so obtained.
• the word intonation database must include, according to the above-mentioned embodiment, at least sixty-six word intonation curves corresponding to twenty-two classes times three positions; the amount of memory required for storing such database is clearly small, at least if compared with the amount of memory necessary for storing the vocabulary.
• the number of predetermined positions may be increased; it is surely useful to consider as a possible position the last but one in the sentence: in fact the last but one word of a sentence, from the intonation point of view, is the preparation to end of the sentence; in addition another possible position is the second in the sentence, particularly useful when the first word in the sentence is short.
• the number of word intonation curves, and consequently of memory required increases respectively to eighty-eight and one hundred and ten.
• the message intonation curve depends, in addition, on the general intonation of the text message, i.e. interrogative, affirmative, exclamatory, dubitative, etc.. If the implementation of the message intonation curve shall take this element into account the number of curves in the database shall increase further: if, for example, three different general intonations are considered, the number of curves could be 198 [3 X 66].
• step a) the value of a fourth parameter is determined corresponding to the general intonation of the text message, and in step b) the word intonation curve corresponds to the values of four parameters.
• a first word intonation curve corresponding to a first word having a number of syllables bigger than the predetermined limit is obtained through the concatenation of at least a second and a third word intonation curves respectively corresponding to at least a second and a third words; the global number of syllables of said second and said third words being equal to the number of syllables of said first word.
• the values of the other two parameters, the position of the accented syllable and the position of the word, should be determined as if the word would be actually splitted into two words: if the position of the first word (9 syllables) is "begin of sentence", the position of second word (4 syllables) is "begin of sentence” and of the third word (5 syllables) is "middle of sentence”; if the accented syllable of the first word (8 syllables) is the seventh one, the accented syllable of the second word (4 syllables) is chosen by chance and the accented syllable of the third word (5 syllables) is the third.
• the method, according to the present invention, of synthetizing an intonated vocal message corresponding to a text message consisting of the concatenation of words includes at first the generation of the pronunciation of the vocal message, at second the generation of the intonation of the vocal message independent and separate from the generation of its pronunciation, and at last the association of the pronunciation and the intonation into the intonated vocal message.
• the intonation of the vocal message may be generated according to a message intonation curve obtained through one of the above-described methods; the pronunciation of the vocal message may be generated either algorithmically or by using a word vocal database.
• a method of synthetizing an "intonated" (having a correct intonation) vocal message corresponding to a text message consisting of the concatenation of words using a word vocal database including for each word a sequence of frames, and using a word intonation database including word intonation curves, each of said word intonation curves being a sequence of pitch values corresponding to a sequence of frames of a word of said word vocal database, and being identified by the value of at least one parameter corresponding the position of such word inside a sentence
• the vocal message is obtained through the extraction of the sequences of frames corresponding to the words of said text message from said word vocal database and the concatenation of said sequences
• the message intonation curve of said vocal message is obtained through the above-described first method
• the intonated vocal message is obtained associating to each frame of said vocal message a corresponding pitch value of said message intonation curve.
• the word intonation curve shall be implemented splitting the group of words into the single words; for example, if the group "The flight for" has to be synthesized, at first the correspondent sequence of frames has to be extracted from the word vocal database, at second three suitable intonation curves corresponding to the words "The", “flight”, "for” has to be extracted from the word intonation database, and at last such three curves has to be concatenated to form a "long” word intonation curve; this requires that the data fields associated to each sequence of frames provide the following information: number of words, numbers of syllables of the different words, and positions of the accented syllables inside the different words.
• the lengths of the various sequences of frames of the word vocal database are quite different and depends at least on the number of syllables of the correspondent various words; in addition such lengths are related to the words themselves.
• such lengths be normalized to predetermined lengths: for example for 1 syllable words a sequence of 15 frames of 20 mS each, for 2 syllables words 30 frames, for 3 syllables words 45 frames, etc..
• the lengths of the various sequences of pitch values are important because they contribute in determining the naturalness of the emitted vocal message. Therefore, during the construction of the word intonation database, they must be carefully considered and, during the synthesis, they must be used as a reference.
• the length of the sequence of frames of a word and the length of the relevant sequence of pitch values should be the same.
• This may be accomplished through the adjustment of the vocal message, obtained by simple extraction and concatenation, in such a way that the length of each of the sequences of frames, corresponding to the words of the text message, be equal to the length of the relevant word intonation curve extracted.
• the adjusting may consist in duplicating or deleting frames and is carried out in such a way as to obtain a uniform expansion or reduction and that the stationary vowel frames of the adjusted sequence of frames respectively coincide to the stationary vowel pitch values.
• the text message includes the Italian word "volo" in second position, that in the word vocal database it corresponds to the following sequence of 26 frames: and that the stationary vowel frames are in position number 10 and 24; let us also suppose that in the word intonation database the intonation curve for a two-syllable word in second position and with the first accented syllable is derived from the pronunciation of the Italian word "molo”, corresponding to the following sequence of 32 frames: having the stationary vowels frames in positions 12 and 28, and is therefore composed of a sequence of 32 pitch values.
• the construction of the word intonation database may comprise the recording of a number of sentences equal to the number of word intonation curves of the database, the extraction of a word from each sentence, and the computation of the pitch value in each frame of the word using a standard technique. If the number of classes considered is 22 and the position considered is 3 the number of sentences is 66.
• the word to be extracted should not be a verb or an adjective but a substantive, should be preceded by a word terminating with a consonant and be followed by a word starting with a plosive consonant. This means that the sentence to be chosen should permit the easy extraction of the intonation curve.
• stationary vowel frames should be identified (one every syllable) and consequently the corresponding stationary vowel pitch values.
• the construction of the word vocal database may comprise the recording of a number of sentences equal to the number of words of the necessary vocabulary, the extraction of a word from each sentence, the classification according to the values of the already mentioned three parameters, the generation of the correspondent sequence of frames, the identification of the stationary vowels frames.
• the speech synthesis system according to the present invention shall synthesize intonated vocal messages using the speech synthesis method described above.
• It comprises the normal HW elements of a standard speech synthesis system but is programmed in such a way as to carry out such method; therefore it may contain, in a program memory, a suitable program and, in a data memory, at least a word intonation database of the kind described above.
• it may comprise further at least a word vocal database of the kind described above.

Landscapes

• Engineering & Computer Science (AREA)
• Acoustics & Sound (AREA)
• Audiology, Speech & Language Pathology (AREA)
• Human Computer Interaction (AREA)
• Physics & Mathematics (AREA)
• Computational Linguistics (AREA)
• Multimedia (AREA)
• Health & Medical Sciences (AREA)
• Compression, Expansion, Code Conversion, And Decoders (AREA)
• Telephonic Communication Services (AREA)
• Machine Translation (AREA)
• Document Processing Apparatus (AREA)
• Reduction Or Emphasis Of Bandwidth Of Signals (AREA)
• Mobile Radio Communication Systems (AREA)
• Electrically Operated Instructional Devices (AREA)

Applications Claiming Priority (2)

Publications (3)

Family

ID=11364645

Family Applications (1)

Country Status (5)

Cited By (1)

Families Citing this family (1)

Family Cites Families (3)

• 1993
  • 1993-01-15 IT ITMI930044A patent/IT1263756B/it active IP Right Grant
  • 1993-04-24 ES ES93106692T patent/ES2149789T3/es not_active Expired - Lifetime
  • 1993-04-24 EP EP93106692A patent/EP0606520B1/de not_active Expired - Lifetime
  • 1993-04-24 AT AT93106692T patent/ATE196210T1/de not_active IP Right Cessation
  • 1993-04-24 DE DE69329375T patent/DE69329375T2/de not_active Expired - Lifetime

Also Published As

Similar Documents

Legal Events