Background of the Invention
This invention relates to speech synthesis of freely generated text and, more particularly, to synthesis of freely generated text created by sound-impaired (hearing-impaired or speech-impaired) persons.
Dual Party Relay Service is a service that allows sound-impaired individuals to communicate with individuals who are not so impaired. By dialing a specified number, such individuals are interconnected to an attendant that connects the calling party to the called party and relays the established conversation.
More specifically, incoming calls from sound-impaired users are placed through a digital device, such as a Telecommunications Device for the Deaf (TDD) or a computer terminal and a modem. The sound-impaired user enters text through the TDD (or the like) and the text is received and displayed to the attendant. Through a separate connection, the attendant speaks the displayed text words to the unimpaired party, listens to the spoken response of the unimpaired party and transmits a written text message to the sound-impaired party which corresponds to the spoken response heard by the attendant. The sound-impaired party receives the written text message on a display device.
The above-described procedure suffers from lack of privacy. It is also very slow, cumbersome, and highly labor intensive. It would be highly beneficial to both the users and the telephone company to eliminate or at least reduce the attendant's involvement in calls to and from sound-impaired parties.
The potential clearly exists for eliminating the need for an attendant's assistance at least in the direction where text is converted to speech. One example is the "telephone information" service of AT&T and the Bell Operating Companies. In response to an inquiry, the telephone information operator accesses a data base, identifies the desired number and activates a synthesizer. However, this and all of the known other text-to-speech synthesizers systems deal exclusively with "canned" speech, which is speech (or text) with a predefined syntax and vocabulary.
In principle, it should be possible to enter the text generated by a TDD user into a speech synthesizer and have the synthesizer generate the spoken words which could be transmitted to a person who is not sound-impaired. In practice, this is not easy to do. Speech synthesizers convert text to speech exactly as the text appears. This requires the text to be correct, error free, properly punctuated and in standard syntax. Written language of the deaf, however, exhibits many departures from expected patterns of English syntax (or other languages, for that matter). The following three sentences are examples of written language of the deaf (WLD) texts that were produced by different writers:
They tend refusing to communicate.
Some people have strong based on knowledges.
I have many appointment from my childrens.
In a study of WLD, V. R. Charrow in Deaf English Technical Report 236, Institute for Mathematical Studies in the Social Sciences, Stanford University, 1974, asserts that the variability of WLD is systematic and rule-governed. She writes:
I suggest that the deaf are constructing a different variety of English, based on an internalization of certain rules of English, but not others, and affected, in some cases, by interference from ASL, whose rules are radically different from those of English. The result is the hypothesized `Deaf English`.
An examination of WLD syntax supports Charrow's claim. Although syntactic variation in WL texts appears random, a closer look shows that the variation is consistent. WLD exhibits a collection of non-standard features that can be identified and described. Some anecdotal evidence for this conclusion comes from the fact that standard English speakers usually adapt to the style after reading a number of WLD texts; serious comprehension problems seem to arise only when sentences are extremely telegraphic, word order is exceptionally free, or lexical choices are obscure.
In short, what is needed for the Dual Party Relay Service is the ability to accept and synthesize freely generated text; and the consistency of WLD suggests that computational analysis of this language style might be a feasible goal.
Summary of the Invention
Motivated by the need to enhance the Dual Party Relay Service, this invention provides for text to speech synthesis of freely generated text. More particularly, this invention offers a text-to-speech synthesizer that accepts a freely generated text sequence of words and synthesizes the sequence with proper emphasis and with properly placed pauses. In combination with other elements, this invention provides for an enhanced Dual Party Relay Service where the text generated by the sound-impaired party is synthesized without the attendant's intervention. The synthesis can take place at the sound-impaired person's premises, at the local telephone central office, or at some designated regional location.
In combination with still additional elements, this invention provides for a completely automatic two-way communication with sound-impaired parties by including a speech-to-text converter.
In accordance with the principles of this invention, text that is generated by a sound-impaired user is made more intelligible by interpreting abbreviations, correcting errors (misspellings and "noise"), translating special terms that are commonly used by sound-impaired persons, translating non-standard word sequences or phrases, and in general, converting the text messages to more standard English. In addition, pauses are inserted to enhance intelligibility.
In one preferred embodiment of this invention, text is inserted into a processor which analyzes a selected number of words at a time. Based on predefined rules, signals are applied to a conventional synthesizer to develop the speech signal.
Brief Description of the Drawing
FIG. 1 depicts a block diagram of a text to speech sound conversion system;
FIG. 2 illustrates the use of the text to speech conversion system in a telephone network, together with a speech to text conversion system;
FIG. 3 presents a flow chart for one manner of signal processing in the system of FIG. 1; and
FIG. 4 presents a flow chart for another manner of signal processing in the system of FIG. 1.
Detailed Description
FIG. 1 illustrates, in block diagram form, one embodiment for the freely generated text-to-speech synthesizer of this invention. Block 10 is an intelligent parser, and block 20 is a synthesizer. Synthesizer 20 is a conventional synthesizer such as the "Olive-Liberman Synthesizer", J. P. Olive and M. Y. Liberman, 1985, Text-to-speech--An Overview, Journal of the Acoustic Society of America, Supplement 1 78, S6. Its input, on line 21, is text that is at times annotated with special instructions. Its output, on line 22, is the analog speech signal.
Intelligent synthesizer 10 comprises word detector 11, "word store" blocks 12, 13 and 14 and rule based processor 15. Text is entered into term detector 11 where words, or terms, are identified. Typically, a blank space, a punctuation mark, or the end of a line denotes the end of a term. The punctuation marks are terms in and of themselves. Detected terms are inserted into the serial interconnection of "word store" 12, "word store" 13 and "word store" 14 in a shift register fashion. Although FIG. 1 depicts the use of three "word store"s (12, 13 and 14), it should be appreciated that any number of such stores may be used. The terms stored in each one of the "word store"s is received by rule-based processor 15.
The function of processor 15 is to analyze the terms applied thereto, to modify the applied terms, to incorporate emphasis and pauses in syntactically proper places, and to apply the modified terms to synthesizer 20 at the appropriate times.
The rules executed within processor 15 can be divided into two classes: the rules that identify and alter terms and the rules that recognize relevant syntactic information.
The rules that identify and alter terms take care of recognizing abbreviations, recognizing special terms that are used by the relevant community of users, recognizing word phrases containing a number of words equal to the number of "word stores" (in FIG. 1, that number is 3) and recognizing misspellings, typographical errors and other "noise" entries. A common "noise" entry, for example, is "XXXXX", which typically denotes deletion of letters (in this case, 5 letters).
In addition to the above word modifications, the rules that identify and alter terms also perform a disambiguation function to make the text more accurate and more intelligible. That includes adding words (e.g. prepositions) to make the text have a more standard English syntax.
After the rules that identify and alter terms are applied, the syntactic rules are applied. These rules apply local syntactic constraints to determine whether a pause should come before or after any word in the "word stores".
In connection with pauses, the issue of input timing and processing timing must be addressed. Timing of spoken speech is quite important to intelligibility or, more accurately, the relative timing of groups of words is critical. However, information about timing is not available in the text supplied by a TDD, and the arrival times of text words bear no relationship whatsoever to the proper timing of the spoken speech. Additionally, the processing times within processor 15 are not known a priori, so the timing of the output signals of processor 15 does not track the timing of the input signals. Furthermore, one can have the situation where closely timed incoming text requires so much processing time that an input buffer must be provided, whereas in other situations, the text may be slow in coming and there is a substantial hiatus in the processing needed of processor 15. Neither situation should be allowed to affect the timing correctness of the synthesized output.
To that end, intelligent parser 10 includes a FIFO memory 16 interposed between term detector 11 and "word store" 12, and a FIFO memory 17 interposed between processor 15 and synthesizer 10. FIFO memory 16 operates under control of term detector 11 and processor 15. As soon as a term is detected, it is stored in FIFO 16; and as soon as the processing of the term in "word store" 14 is completed, the oldest term in FIFO 16 is injected into "word store" 12.
FIFO 17 is also under control of processor 15, but it works in a slightly different mode. As soon as a term is modified by processor 15 it is inserted into FIFO 17. Modified terms continue to be inserted until processor 15 determines that a pause in the speech is called for. At that time, all of the terms sorted in FIFO 17 are extracted, in order, and sent to synthesizer 20. In this manner, the completely arbitrary timing of the incoming words is replaced with a timing that comports with the rhythm of natural speech.
The structure of FIG. 1 contemplates different and distinct hardware for carrying out the functions for the intelligent parser-synthesizer of this invention. Such a parser-synthesizer may be placed in each local telephone central office, in some regional office, or associated with a particular type of service (such as the called-party-paying "800 service" or the calling-party-paying "900 service").
Such hardware may also be realized in a very compact arrangement that is embodied in one or very few interated circuits. When that realization is carried out, and when the price is low enough, the intelligent parser-synthesizer of FIG. 1 may be included in a customer premises apparatus. FIG. 2 depicts such an arrangement, which includes a subsystem at the customer's premises that is connected to the telephone network. The subsystem of FIG. 2 includes the intelligent parser-synthesizer for the outgoing path, and a speech-to-text converter 30 for the incoming path. Both paths are connected to the telephone network, and through that network the subscriber at the left portion of the drawing is connected to some other subscriber at the right portion of the drawing.
Speech-to-text converters 30 are not as yet robust enough to recognize and detect free flowing speech of arbitrary speakers, but some systems are available that can be trained to recognize up to 5000 words, when trained to a particular speaker. Other systems recognize a more limited vocabulary but from any speaker. Even this modest capability is useful, because the sound-impaired user would not need to enlist the assistance of an attendant at all in some instances, such as when "speaking" by telephone with particular persons (e.g. spouse) when a 5000 word recognition system is used.
There is another aspect to the form of the hardware embodiment that the intelligent parser-synthesizer may take. That is, the FIG. 1 structure may be implemented with a program controlled processor that encompasses all of the elements in FIG. 1, or at least all of the elements within block 10. FIG. 3 presents a flow diagram of the processing performed in a hardware embodiment of FIG. 1, while FIG. 4 presents a flow diagram of the processing performed with a program controlled processor that encompasses all of block 10.
The first block in FIG. 3 (block 100) converts each term that enters "word store" 12 into a data structure. That is, each term is made a part of an associated set of information items, such as the term's spelling, the part of speech to which the term belongs, whether the term is an abbreviation, etc. (Since the terms stored in "word stores" 13 and 14 had previously been in "word store" 12, their data structures are already stored in processor 15.)
With the data structures established for the terms in the "word stores", the next step (block 110) is to apply the word adjustment rules. The consequences of the word adjustment rules are stored in the data structures. Next, in block 120, the syntactic rules are applied to determine whether pause flags or a de-emphasis flag should be set. The consequences of this analysis are also stored in the data structures. Finally, in block 130, the processed term of "word store" 14 is stored in FIFO 17.
Block 140 executes the query that leads to an output being provided to the synthesizer. When it is determined that the term stored in FIFO 17 includes a "pause-before" flag, then the FIFO is emptied (the stored information is sent to the synthesizer) and then the term is stored in the FIFO. When a "pause-after" flag is found, then the term is first stored and after the storing action the FIFO is emptied. When neither flag is found, the term is merely stored. In any event, as depicted in FIG. 3, the process continues to block 100 to receive the next term from FIFO 16, as it is loaded into "word store" 12.
FIG. 4 depicts the process by which the intelligent parser of this invention is realized with a program controlled processor. Block 200 stores the input text string into memory and sets a variable *terminal-string* to be the input text string. Block 210 converts each word in *terminal-string* into a "term" data structure with the following fields:
spelling=the spelling of the word, or the printed form of the word, minus any punctuation (? !128). Note that the period (.) is not included as a punctuation because some common TDD words end in a period.
category=if first character of the word is a digit (1 2 3 4 5 6 7 8 9 0), return NUM as category otherwise, the category of the term is the lexical category of the term, which corresponds to the "part of speech" of the known given English or TDD-dialect word. Unknown words are assigned a category of NIL.
special-categories=Some "words" belong to one or more word classes in addition to "parts of speech". This field identifies those special categories (see below).
previous=pointer to previous word in the list that makes up *terminal-string*, if any; else NIL.
next=pointer to next word in the list, if any, else NIL.
pause-before=initially empty (NIL); may later be populated to insert pause by the appropriates character string that is recognized by the synthesizer.
pause-after=initially empty (NIL); may later be populated to insert pause by the appropriates character string that is recognized by the synthesizer.
destress=initially empty (NIL); may later be populated to modify the synthesizer's default speech parameters for this word.
punctuation=contains the punctuation which terminates the word (either;, ! or ?).
The above structure is merely illustrative and other structures can, of course, be employed.
The special categories of a word (third field in the data structure above) may comprise
Returning to FIG. 4, block 220 sets the variable *first-term* to point to the first term in the list created by block 210. This sets the stage for the beginning point where terms will be delivered to the synthesizer. Thereafter, block 230 sets the pause-before field of the first term (pointed to by *first-term*) to "true".
Block 240 begins the actual processing phase by applying rules for identifying misspellings and noise entries. The processing of block 240 is quite conventional in the sense that in the checking the spelling it may borrow from the approaches taken by any one of the many word processing products that are commercially available, such as "word", "word perfect" etc. These products have mechanisms for determining whether a word is found in a dictionary and, in cases where the word is not found, what correctly spelled words most resemble the misspelled word. A best match procedure can also be employed, as described, for example, in U.S. application titled "A Method for Correcting Spelling using Error Frequencies", filed Jun. 14, 1990. Once the correctly spelled word is identified in a table, the category of the word, as well as its special categories, can be extracted from the same table.
Block 250 continues the word alteration processing. By sliding a 3-word "window" along the list of terms in *terminal-string*, it filters out abbreviations and unorthodox contractions, and attempts to disambiguate the text. That is, it attempts to replace the shortened input text with the proper full-spelling word texts that would be synthesized. Block 250 also performs the converse function of contracting some two word and three word expressions to function as single words; at least for purposes of syntactic analysis. Because these rules may change the number of words in the term list, they often perform splicing of the previous and next list, as well as updating the variable *first-term*. (A few of these rules also refer to the term preceding *first-term*). The following are examples of abbreviations, unorthodox contractions, two word phrases and three word phrases.
The rules embodied in block 250 may be customized to the application at hand and modified by the experience gained. The following is one example of such rules, (described in the order applied).
Followikng block 250, block 260 continues the processing by introducing phrase rules that add prosodic cues to the speech synthesizer by filling in the pause-before, pause-after, and destress fields of selected terms. The rules that are carried out may be as follows:
Lastly, block 260 creates an output file for the final pronunciation string. It processes the list of terms from first to last, as follows: