Synthesis-based speech training system and method

Abstract

This invention includes a speech training system that allows a student to enter any utterance to be learned and have the articulatory model movements required to produce the utterance displayed on a CRT screen. The system accepts a typed utterance, breaking it down into a set of speech units which could be phonemes or syllables and the onset and offset of the speech units. The set of speech units is sent to a synthesizer, which produces a set of parameters indicating the acoustic characteristics of the utterance. The acoustic parameters are converted into articulatory parameters emphasizing the frequency and nasality required to produce the typed utterance. The speech units and onset and offset of each are used to generate tongue-palate contact patterns required to produce the typed utterance. The articulatory parameters are displayed on the CRT screen. The acoustic parameters are also sent to a formant synthesizer which converts the parameters into speech output. The system measures a student's production and then evaluates the student's production against the parameters of the typed utterance for its similarity. Feedback on the similarity is displayed on the CRT screen.

Description

BACKGROUND OF THE INVENTION

1. Field of the Invention

This invention relates to speech training systems and, more specifically, to a speech training system which permits a speech-impaired student to type in any word or sentence he or she wants to learn and observe, on a CRT display, the articulatory model movements required to produce that word or sentence in the form of tongue-palate contact patterns. The system is particularly suited to deaf children, who do not receive auditory information and who tend to learn to type at an early age. The invention can also be used to help normal-hearing students learn how to speak foreign languages.

2. Description of Related Art

The most fundamental method for teaching deaf children to speak consists of teachers using their own vocal apparatus for showing the correct vocal gestures. Children can observe the external appearance of the lips, jaw and, to a limited extent, the tongue, as speech is produced by a teacher. Children are sometimes instructed to use tactile feedback from the teacher's vocal organs for comparison to their own. This method has obvious limitations in that many of the articulatory gestures in speech are not observable externally.

In recent years, it has been possible for teachers to demonstrate how speech is produced with the aid of instruments and of computer programs which analyze speech. These instruments and programs permit the observation of many characteristics of speech, including its acoustic manifestation, as the speech is being produced. This system is best demonstrated in the Computer Integrated Speech Training Aid (CISTA) developed by Matsushita. The CISTA provides multiple-channel data gathered by several transducers.

1. The dynamic palatograph. This instrument, whose use was first reported by 1962 by a Soviet researcher, Y. Kuzmin, indicates contact between the tongue and palate by means of a number of electrodes on an artificial palate that is worn in the mouth. When the tongue touches one of the electrodes, a low-voltage circuit is completed which is registered by instruments outside the mouth. The indication as to the presence or absence of contact is provided on a CRT display.

2. Nasal sensor. An electret microphone held to the side of one nostril by headgear or temporarily attached with adhesive tape provides an indication of nasal vibration.

3. Throat sensor. An electret microphone, held at the larynx by means of a flexible collar, provides an indication of glottal vibration.

4. Airflow sensor. Several methods have been used for sensing airflow, using a device held by the child in front of the mouth.

5. A standard microphone provides input for acoustic analysis.

The teaching of speech to the deaf is made difficult by the limited time the children have with speech teachers. While a hearing child receives speech input and acoustic feedback as to his own production for many hours a day, a deaf child typically only receives such feedback during training sessions which may be as infrequent as one session a week.

Combining such speech training devices as the instruments of the CISTA with computers allows children to receive part of their training without a teacher's intervention, greatly expanding the time that training is available. However, speech training devices such as CISTA which provide feedback directly to children are limited to teaching individual sounds or a limited set of preprogrammed utterances. The purpose of the present invention is to permit a child to receive information as to the production of any utterance without the assistance of a teacher.

Other prior art text-to-speech systems permit any utterance which is typed to be automatically synthesized. A device called "DECTalk," produced by Digital Equipment Corporation, is currently the best-known example of text-to-speech for English. All these text-to-speech systems are limited to producing audible sounds only.

The shape of the human vocal tract determines the resonances that in turn control human speech output. Electronic and computational models of the relationship between the vocal tract shape and the resulting acoustic output have been an important part of speech research for many years. In this work, the shape of the vocal tract was provided by the researchers, and the acoustic output was measured.

More recently, researchers have been developing articulatory synthesis, in which the generation of the vocal tract shapes was produced automatically. In this case, the input consists of a string of phonemes. This string is converted to a string of vocal tract shape specifications. The vocal tract shapes are then used in a vocal tract model which produces speech.

The automatic generation of vocal tract shapes from phonemes in this earlier work did not include the generation of tongue-palate contact patterns. Use of synthesized tongue-palate contact patterns to teach speaking has not been contemplated heretofore.

OBJECTS AND SUMMARY OF THE INVENTION

An object of the present invention is to provide a speech training system that permits a student to type into a computer any utterance desired to be learned and have the articulatory model in the form of tongue-palate patterns displayed on a CRT screen.

Another object of this invention is to permit a student to attempt to speak the typed utterance and receive feedback, on the CRT screen, on the similarity of his or her efforts to the model production displayed.

This is accomplished by the system first editing the typed utterance and sending it to a synthesis unit. The synthesis unit parses the utterance, analyzes it into a phonemic string, and creates a set of parameters that specify the acoustic characteristics of the utterance for every 10 ms. The parameters are sent back to the apparatus and are converted to a set of parameters that indicate the articulatory model, the tongue-palate patterns required to produce the typed utterance. The tongue-palate patterns are then displayed on a CRT screen.

By using a set of transducers to measure the student's nasal vibration, glottal vibration, expiratory airflow, tongue contact, and audible sound, the speech training system evaluates the student's enunciation for its similarity to the model tongue-palate contacts. The evaluation is then displayed on a CRT screen.

BRIEF DESCRIPTION OF THE DRAWINGS

The exact nature of this invention, as well as its objects and advantages, will become readily apparent upon reference to the following detailed description when considered in conjunction with the accompanying drawings, in which like reference numerals designate like parts throughout the figures thereof, and wherein:

FIG. 1 is a block diagram showing the steps involved in converting an ASCII orthographic representation for a word into phonemes, stress, and parts-of-speech information.

FIG. 2 is a block diagram showing the speech training system;

FIG. 3 is a sound spectrum showing the acoustic parameters of the nasal zero and nasal pole produced during an utterance;

FIG. 4 is a sound spectrum showing the parameters of tongue contact pattern against the palate during an utterance; and

FIG. 5 is a display of an actual tongue contact pattern created during an utterance.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

The following description is provided to enable any person skilled in the art to make and use the invention and sets forth the best modes contemplated by the inventor of carrying out his invention. Various modifications, however, will remain readily apparent to those skilled in the art, since the generic principles of the present invention have been defined herein specifically to provide a synthesis-based speech training system constructed in accordance with the teachings of the invention shown in FIGS. 1-5 of the drawings, in a manner to be described in greater detail below.

The synthesis-based speech training system can be implemented on a laboratory digital computer. The system consists of a means for entering an utterance to be learned, which is preferably a keyboard 1 connected to the speech training system (FIG. 2). The entered utterance is transferred to the speech training system in ASCII representation. The system then uses a text-to-speech system to automatically synthesize the typed utterance. A device called "DECTalk," produced by Digital Equipment Corporation, is currently the best-known example of text-to-speech for English. A newer text-to-speech system, STLtalk, is being developed at Speech Technology Laboratory, division of Panasonic Technologies, Inc., and is actually being implemented in the invention.

Communication between the synthesis unit 3 and the speech training system is done via an RS-232 port. The RS-232 configuration is set for XON/XOFF protocol and the transfer settings are 9600 baud, 8 data bits; 1 stop bit, and none for parity. The synthesis unit 3 receives the input text string of phonemes and places it in an input buffer.

The text-to-speech system parses the entered utterance at stage 2. The system performs a syntactic analysis of the utterance based on locations of any orthographic commas, as well as the syntactic role of function words and verbs that are detected during the dictionary matching process.

Next, a phonemic representation is obtained as shown in FIG. 1. Each word is compared with entries in a small pronunciation dictionary 12. If no match is found, the word is broken into small pieces (morphemes) by attempting to remove common suffixes such as "ed" and "ing," etc. at stage 14. Then the remaining root is again compared with entries in a phonemic dictionary 16. If there is still no match, a set of letter-to-phoneme rules is invoked to predict the pronunciation at stage 18. In addition, a part of the phonemic conversion process concerns the derivation of a stress pattern for the syllables of a word. Stress must be predicted if the word is not in the system vocabulary, or if the orthographic word is broken down into root plus affixes and an affix changes the stress pattern given for the root. The stress level of the syllable will be indicated by inserting a stress symbol just prior to the vowel in the phonemic representation. Absence of a stress symbol means that the symbol is unstressed. The affixes stripped at step 14 are reattached at step 20 to provide the phonemic representation.

English language systems utilize the phoneme as the basic speech unit. Other languages, such as Japanese, for example, utilize the syllable as the basic speech unit.

Thus, it should be understood that although the present description of applicant's preferred embodiment describes an English language system based on phonemes as the basic speech unit, the principles stated apply equally to other than English language systems that utilize syllables as the basic speech unit.

The synthesis unit 3 processes this buffer and produces a set of 20 acoustic parameters for every 10 ms of the string of phonemes. The acoustic parameters specify the acoustic characters of the utterance to be demonstrated. The parameters are created by matching the transferred information with predetermined measurements in values based upon the fundamental frequencies and amplitudes, the formant frequencies and bandwidths, and the noise source for speech production of the utterance to be demonstrated. A sample of the acoustic parameters is set forth below.

______________________________________N Symbol Name______________________________________1 AV amplitude of voicing in dB2 AF amplitude of frication in dB3 AH amplitude of aspiration in dB4 AVS amplitude of sinusoidal voicing in dB5 F0 voicing fundamental frequency in Hz6 F1 first formant frequency in Hz7 F2 second formant frequency in Hz8 F3 third formant frequency in Hz9 F4 fourth formant frequency in Hz10 FNZ nasal zero frequency in Hz11 B1 first formant bandwidth in Hz12 B2 second formant bandwidth in Hz13 B3 third formant bandwidth in Hz14 A2 second parallel formant amplitude in dB15 A3 third parallel formant amplitude in dB16 A4 fourth parallel formant amplitude in dB17 A5 fifth parallel formant amplitude in dB18 A6 sixth parallel formant amplitude in dB19 AB bypass path amplitude in dB20 OQ open quotient______________________________________

Two identical sets of the acoustic parameters are created. Following the reception of the utterance termination character, a period, the first set is stored in an array.

The second set of acoustic parameters is then sent on to a formant synthesizer 4 (FIG. 2) which converts the acoustic parameters into an output signal in the form of an analog speech signal. The output of formant synthesizer 4 is then played over a loud speaker 5 for those students who have some residual hearing.

The first set of acoustic parameters which were arrayed in the synthesis unit 3 are then sent back to the speech training system through the RS-232 port. The speech training system reads in these values and stores them into another array. The speech training system then converts the acoustic parameters into articulatory parameters at stage 6. The articulatory parameters show three forms of the typed utterance.

The speech training system converts one group of the acoustic parameters to articulatory parameters showing various attributes of the typed utterance. The group of acoustic parameters used includes the fundamental frequency, the amplitude of the utterance, the presence or absence of voicing, and the frequency of formants 1, 2, and 3. The text-to-speech system supplies the group of acoustic parameters to the speech training system in Hertz multiplied by 10. The speech training system then divides the parameters by 10 before displaying them on the CRT screen.

The speech training system also produces articulatory parameters showing the nasality of the typed utterance. Nasalization in speech is appropriate for sounds such as /m/, /n/, and others, i.e. SIL, AX, AN, DN, ERT, AY, (FIG. 3) and is frequently present in the speech of deaf speakers for sounds where it is inappropriate and impairs communication. Nasalization is characterized acoustically in the frequency domain 22 by the presence of both nasal poles and nasal zeros. Typically only one nasal pole and one nasal zero is acoustically significant. These only appear when the velum, the valve between the nose and the mouth, is in its lowered, open position. It disappears when the velum is closed.

The text-to-speech system supplies a variable of acoustic parameters measuring the nasality of the typed utterance. The speech training system uses the acoustic parameters to create a nasalization index. The text-to-speech system produces acoustic parameters for non-nasal sounds by setting the nasal pole and nasal zero at the same frequency and bandwidth, so that they cancel each other out. The text-to-speech system used in the present invention sets both the nasal pole and nasal zero at 250 Hz for non-nasal sounds. For producing nasal sounds, the frequencies of the nasal pole 28 and nasal zero 26 are moved to different values so that they each have an effect in different parts of the sound spectrum. The frequency of the nasal zeros is temporarily increased to 330 or 360 Hz, depending upon whether the sound is /n/ or /m/ 34, 36. This is shown in FIG. 3. The text-to-speech system also provides the amplitude of the nasal formant 30, 32. Typically, values of the amplitude of the nasal formant range between 30 and 60, (see graph 30) but if it is equal to zero, there is no nasalization 30, 32. Using the acoustic parameters supplied by the text-to-speech system, the speech training system uses an equation to create a nasalization index indicating to the student the amount of nasality required to produce the typed utterance. One form of the equation is: ##EQU1## where:

NI=nasalization index

F.sub.nf =frequency of the nasal formant

F.sub.nz =frequency of the nasal zero

AN=amplitude of the nasal formant

All the input variables are integers

The second part of the right side of the equation is included to multiply the value by zero if the amplitude of the nasalization is zero, and to multiply the value by one if it is not.

The speech training system also creates articulatory parameters showing the tongue-palate contacts required to produce the typed utterance. Besides providing acoustic parameters, the text-to-speech system is used to provide the timing of four points: the onset of a sound, the time when maximum amplitude is achieved, the time when amplitude begins to decay, and the offset. These times are then coordinated to a set of palatographic images. The maximum area of contact for each speech unit, which may be a phoneme or a syllable, is stored, at stage 6, and treated as the target patterns for that speech unit. Since speech sounds vary in contact pattern for different contexts, a number of different contact patterns need to be stored for each context, as well as for each sound.

The durations of the onset, offset, and central steady-state portion are assigned for each speech unit. The contact areas between the tongue and palate are defined by a set of 63 points. A hierarchy is defined for the order in which points make and release contact in the onset and offset portions. Typically, the first points to make contact are those towards the back of the palate, and the first points to release contact are towards the front of the palate.

Since it is difficult to convert the acoustic parameters produced by the text-to-speech system to the tongue-palate contact patterns one must learn, the present invention uses the onset and offset of consonants for synthesizing the tongue-palate contact patterns. For example, in the utterance "She said," the text-to-speech system provides, along with synthesis parameters, the beginning 50 and end 52 of the sound /SH/ and the beginning 54 and end 56 of the sound /s/. This can be seen in FIG. 4, which shows all the parameters for the utterance "She said" including the format bandwidths in Hz 38, 40, 42, and the format frequencies in Hz 44, 46, 48. For palate contact, only the beginning and ending time is used by the present training system, which supplies the actual contact pattern.

The tongue-palate contacts displayed for the student by the present system are contained in the speech training system. A typical pattern is shown in FIG. 5, which shows the tongue-palate contact patterns for vowels surrounding the /s/ in "She said." The larger dots represent contact; the smaller dots represent no contact. Each frame 58 to 104 represents 10 milliseconds.

The frame sequence of FIG. 5 shows the progression from the vowel in "she," followed by the /s/ and the vowel in "said," as a series of 10-millisecond frames 58-104. The first seven frames, 58, 60, 62, 64, 66, 68, and 70, show the contact for the vowel in "she" going into the following /s/ frame 104 of "said." The full /s/ begins at frame 8, 72 and starts to release nine frames later at frame 90. By the 24th frame 104, there is a complete release of contact required for the vowel in "said."

The actual tongue-contact patterns displayed in the teaching program come from three different sources.

The system contains a set of prestored tongue-contact patterns for the consonants /t/, /d/, /s/, /z/, /sh/, /zh/, /l/, /n/, and /r/. At present, the contact for /k/ and /g/ spoken immediately before or after front high vowels are not included, but the principle is the same.

A teacher can put a tongue-contact pattern into the system for any sound in any environment.

A teacher can select a tongue-contact pattern from a production that the child has said well once but needs practice in repeating. This is frequently the most useful sort of pattern, because each person's optimal tongue-palate contact depends strongly on his or her particular palate shape.

The teacher sets which of these sources is used for each consonant. The prestored tongue-contact patterns are used as the default. For all sources, the system provides the contact pattern sequence beginning before the onset and following the release of the consonant.

When an utterance is typed into the system that calls for a consonant with tongue-palate contact, the text-to-speech system tells the training system when that consonant begins and ends. The sequences that have been stored by one of the three methods described above are then used in displaying the proper contact pattern for the student.

After viewing the tongue-palate contact pattern required to produce the typed utterance, the student may then attempt to produce the typed utterance by attempting to speak the utterance. The speech training system uses the instrument of the Computer Integrated Speech Training Aid (CISTA) described above. The instruments of the CISTA use several transducers to measure the student's production. The student's tongue-palate contact, nasal vibration, glottal vibration, expiratory airflow, and acoustic production are entered into the speech training system in the form of measurement by speech transducers 8.

The speech training system then evaluates the measurements of the student's production against the acoustic parameters for the typed utterance for similarity at stage 9. The speech training system then provides the student with feedback on the difference between his or her produced model movements and those model movements required to produce the typed utterance by feedback display 10.

A second embodiment of the invention includes a synthesis-based speech training system to help students with normal hearing to learn to speak foreign languages. The second embodiment uses the same text-to-speech training system as the first embodiment, notwithstanding the following differences.

The student enters text of the target language into the speech training system by keyboard 1. The text-to-speech system edits the entered text at stage 2. The edited text is sent to a synthesis unit 3. The synthesis unit 3 processes the text and produces a set of 20 acoustic parameters. The acoustic parameters specify the acoustic characters of the text in a predetermined foreign language.

Two identical sets of the acoustic parameters are created. The second set is sent on to a formant synthesizer 4 which converts the acoustic parameters into an output signal in the form of an analog speech signal. The formant synthesizer output is then played over a loud speaker 5 in the form of the predetermined foreign language.

The first set of acoustic parameters is then sent to the speech training system. The speech training system then converts the acoustic parameters into articulatory parameters at stage 6 to show the articulatory model movements required to produced the entered text in a predetermined foreign language. The articulatory parameters, including frames of tongue-palate contact patterns for consonants, are then displayed on a CRT screen 7, or similar display.

After viewing the articulatory model movements, including the tongue-palate contact patterns required to produce the entered text in the foreign language, the student may attempt to produce the text in the foreign language by speaking the entered text in the foreign language. The speech training system measures the student's speech by transducer 8. The speech training system then evaluates the measurements of the student's production against the acoustic parameters for the entered text in the foreign language at stage 9. The speech training system then provides the student with feedback on the difference between his or her produced model movements and those model movements required to produce the entered text in the foreign language on a feedback display 10.

Those skilled in the art will appreciate that various adaptations and modifications of the just-described preferred embodiment can be configured without departing from the scope and spirit of the invention. Therefore, it is to be understood that, within the scope of the appended claims, the invention may be practiced other than as specifically described herein.

Claims

1. A synthesis-based speech training apparatus, comprising:

input means for entering an utterance;

editing means for converting the utterance into a set of speech units and the onset and offset of each speech unit;

synthesizer means for converting said set of speech units and onset and offset of speech units into a set of tongue-palate contact patterns arranged in a hierarchical order; and

means for displaying the set of tongue-palate contact patterns for the entered utterance.

2. The speech training apparatus of claim 1 wherein said synthesizer means provides the durations of the onset, offset, and contact steady-state portion of each speech unit.

3. The speech training apparatus of claim 2 wherein the times of onset, offset, and steady-state are assigned for each speech unit by said synthesizer means.

4. The speech timing apparatus of claim 3 wherein said synthesizer means stores the maximum area of tongue-palate contact for each speech unit along with at least one alternate contact area for each different context in which that speech unit is used.

5. The speech training apparatus of claim 4 wherein each contact area between the tongue and palate is defined by a set of 63 points.

6. The speech training apparatus of claim 1 further comprising means for measuring acoustic parameters of a spoken utterance, including measuring the tongue-palate contact pattern for the utterance;

means for evaluating the similarity between the utterance being demonstrated and the measured acoustic parameters and measured tongue-palate contact pattern; and

display means for displaying the differences between the utterance being demonstrated and the measured utterance.

7. A synthesis-based speech training method comprising the steps of:

determining an utterance to be learned;

converting said utterance into a speech unit and the onset and offset of the speech unit;

converting said speech unit and onset and offset into a set of tongue-palate contacts; arranged in a hierarchical order; and

displaying the set of tongue-palate patterns for the utterance to be learned.

8. The speech training method of claim 7 wherein said speech unit converting step comprises:

storing an area of tongue-palate contact for each speech unit; and

assigning times of onset, offset, and steady-state for each speech unit.

9. The speech training method of claim 8 wherein said storing an area of tongue-palate contact step comprises:

storing the maximum area of tongue-palate contact along with at least one alternate contact area for each different context in which the speech unit is used.

10. The speech training method of claim 9 wherein said storing step comprises:

storing each contact area between the tongue and palate by a specific set of 63 points.