Search

A wide range of speech investigation requires natural-sounding synthetic speech whose individual acoustic features-such as pitch, formants, and duration-can be varied independently. This study describes an algorithm for obtaining such synthetic speech by automatically analyzing natural speech to obtain data for the control of a formant synthesizer. The lowest three formants of natural speech were determined by finding simultaneous solutions of the least-square-fit equations by means of a Newton-Raphson technique. The data obtained from this formant analysis, together with pitch data which were also obtained automatically, were used for controlling a computer-simulated formant synthesizer. This algorithm was tested for various sentences and speakers, and the synthesized speech was found to be of good quality. [ABSTRACT FROM AUTHOR]

In speech synthesis, utterances are created by concatenating elements from an acoustic inventory made up of phonemes, dyads, syllables, or other convenient units. For further development of the AT&T text-to-speech system [J. P. Olive and M. Y. Liberman, J. Acoust. Soc. Am. Suppl. 1 78, S6 (1985)] we need a fast, simple method to identify problems with elements in the acoustic inventory. Phonetic transcription is one method that has been used in the past for this type of synthetic-speech evaluation. However, transcription experiments are difficult to run and require specially trained listeners. Other methods that have been used, e.g., the diagnostic rhyme test, are better suited to provide comparative information about several synthesizers than the diagnostic information needed for this evaluation. To fill this gap, we have developed a procedure that we call verification. It is an interactive, computer-based procedure that can be used with listeners who have little training. In this talk we describe the procedure and its early application. [ABSTRACT FROM AUTHOR]

We discuss the design of a text-to-speech synthesizer, which accepts any type of English text as input, and creates an appropriate speech signal as output. Effective algorithms for converting text to sound must make use of intermediate data structures that systematically encode the degrees of freedom available to speakers of the language being synthesized. These data structures are an engineering approximation to what linguists call phonological representations; we will call them 'P-structures.' Any TTS system must: (1) define its version of P-structures; (2) design and implement algorithms for transforming input text into P-structures; (3) design and implement algorithms for transforming P-structures into sound. Task [sharp/mesh/hash_synonym]1 is mainly a problem in applied linguistics; task [sharp/mesh/hash_synonym]2 can best be seen as applied AI; task [sharp/mesh/hash_synonym]3 is an application of phonetics and signal processing expertise. Because of the interdependence of approximate solutions in different portions of the system, integration of the various parts is a non-trivial problem. We will analyze an example of a TTS system, showing how these problems were handled in building and combining its numerous pieces. Finally, we point out some areas where better solutions are needed, and suggest how to find such solutions. [ABSTRACT FROM AUTHOR]

Digital computers have been used to process and generate various acoustic signals during the past two decades. Because of needs for real-time interaction with computers, we first introduced the use of hybrid systems, where the computer controlled analog sound generating equipment. More recently, a great deal of special purpose hardware has been built to generate sounds digitally in real-time. Recent advances in VLSI design have brought a new generation of components called the Digital Signal Processor Chips. These chips are relatively inexpensive, and are capable of doing many of the complex operations required for the synthesis of signals such as speech, tones, and music. This paper will desribe a method for using such DSP chips in a real-time scheme. Such a scheme is presently used in speech analysis and synthesis systems, and could be adapted for music and tone-signaling applications. [ABSTRACT FROM AUTHOR]

A speech spectrum is a product of the source spectrum and the vocal tract shape spectrum. Most analysis techniqus do not differentiate between the source and the vocal tract spectrum. The poles derived from an LPC analysis of the speech signal describe the influence of both source and vocal tract functions. This paper will describe a method to separate the LPC derived poles into two groups describing the source and vocal tract spectrum. The major criterion used in the algorithm was the spectral flatness of the source. In order to evaluate this method, the vocal tract poles were used to control a series formant synthesizer driven by a fixed source. [ABSTRACT FROM AUTHOR]

This paper describes a procedure for determining LPC area parameters in real time. The analysis involves the following steps: High-frequency pre-emphasis, windowing, computing the autocorrelation coefficients, and their conversion to area parameters directly from the autocorrelation coefficients without first computing the predictor coefficients. Computer simulations show that 20-bit precision is sufficient to avoid significant errors in the conversion of the autocorrelation coefficients to area parameters if frame-by-frame normalization is used. Our results show that good speech quality can be maintained in synthesis by using 16 area parameters updated at 10-ms intervals. The real-time requirement of this analysis procedure cannot be met by any single signal processor chip which is currently available. We have, therefore, used a multiprocessor configuration of signal processor chips to perform the analysis in real-time. The Bell Laboratories DSP chip-capable of performing a 16 by 20 bit multiply with 40 bit accumulation in 0.8 μs-was found to provide the necessary precision and speed. [ABSTRACT FROM AUTHOR]

This paper describes a small real-time speech synthesizer. The synthesizer requires as its input a string of phonemes and the associated duration, pitch, and amplitude parameters. The synthesis scheme uses a large table of stored transitions (dyads) between phonemes. These transitions are stored in terms of LPC-derived area parameters. To obtain the area parameters for a particular string of phonemes, the appropriate transitions are retrieved and their durations adjusted to reflect the desired durations of phonemes in the string. The speech synthesizer consists of an LS111/23 microcomputer, a digital signal processor chip, and a digital-to-analog converter. The microcomputer processes the input, retrieves the stored transitions, adjusts the durations, and smooths the area, pitch, and amplitude parameters across the dyadic boundaries. The smoothed parameters are used to synthesize the speech signal on the digital signal processor. [ABSTRACT FROM AUTHOR]

Traditional metrical notation specifies at least two kinds of information - (1) nominal point of occurrence for each note and (2) a grouping of notes with implication for their relative stress, often in several levels of subdivisions. These two kinds of information are not the same; for example, the same set of six numerically equidistant time points can be grouped as 2 threes or 3 twos. For performance, the nominal duration of a note is presumably obvious (e.g., a 16th note at d=60 MM is, in principle, 250 msec). The objective interpretation of the metric grouping of notes is not at all obvious (e.g., how do we distinguish 2 threes from 3 twos?). We have studied this question both by analysis of live production and by computer synthesis, concentrating on patterns in which the same sequence of nominal durations is subject to different metrical groupings. Listening to synthetic patterns reveals that even small differences in onset-to-onset duration, articulative space, on time, timbre, or attack can strongly affect the way we group notes perceptually. We will discuss the use of such differences in live performance, giving examples. [ABSTRACT FROM AUTHOR]

Speech can be synthesized by concatenating dyads representing the transition between phoneme pairs. The resulting speech should be well coarticulated if the dyads themselves are free of coarticulatory influences and are therefore independent of context. We tested dyads for anticipatory coarticulation in 151 words as follows. Within each word, the /XæY/ sequence, where /X/ and /Y/ were consonants. were dyadically synthesized. The word /pæn/, for example, was synthesized with the dyads /pæ/ and /æn/. The /pæ/ dyad was an archetypal form derived from many words (pan, papoose, pad, sphagnum, etc.); the /æn/ dyad was taken unchanged from /pæn/. Only the segmental features were synthesized; the prosodic features were left intact. Listeners noticed degradation in words where anticipatory coarticulation of a nasal /Y/ consonant caused /æ/ to be nasalized: words with non-nasal /Y/ were well synthesized. We conclude that nasalized vowels are needed in dyadic synthesis to cope with anticipatory coarticulation of nasal consonants. [ABSTRACT FROM AUTHOR]

A hard-wired real-time linear predictive coding (LPC) synthesizer has been built. The synthesizer computes the speech signal from data which it receives pitch synchronously from a computer: the data includes a pitch period, sum of squares amplitude, and 12 LPC coefficients. In computing the speech signal for one pitch period, the overhang energy from previous pitch periods contributes to the energy of the signal; this factor must be taken into account in order to eliminate roughness in the resultant speech. The device computes a scaling factor for the excitation function which allows for this overhang energy. This requires substantial additional computation; therefore, fast arithmetic processors were designed so that speech can be synthesized in real time. All computation is performed in floating-point arithmetic to eliminate round-off errors. The resulting device can synthesize speech in real time, and the speech quality is the same as that of the software version.11 [ABSTRACT FROM AUTHOR]

A speech synthesis model in which the formant resonators are connected in series does not produce satisfactory speech quality for sounds produced with a narrow constriction in the vocal tract or with the velum lowered so that the nasal cavity is coupled to the oral cavity. These latter cases are better handled by a parallel connection of formants, in which the amplitudes of the individual formants can be controlled independently. An existing analysis-synthesis method for a series model of speech [J.P. Olive, 'Automatic Formant Tracking by a Newton-Raphson Technique,' J. Acoust. Soc. Am. 50, 661-670 (1971)] has been adapted to a parallel formant model of speech. The analysis procedure derives pitch and four formant frequencies and amplitudes from natural speech which are then used to control a speech synthesizer. The quality of synthetic speech produced in this manner is a substantial improvement over that produced according to a series model. [ABSTRACT FROM AUTHOR]

The present study investigated whether the prosodic cue (duration) serves to disambiguate syntactically ambiguous sentences. An example of one of the ten sentences used is, 'The hostess greeted the girl with a smile.' Some of the sentences had two alternative bracketings, e.g.: 'The hostess [[greeted] [the girl]] [with a smile],' and 'The hostess [greeted] [[the girl] [with a smile]],' while others did not. In this experiment, the duration of certain key words was systematically varied. The groups of words that were varied in this example were: 'greeted the' and 'girl with a'. The test sentences were presented to subjects who chose between the two alternative meanings. The results indicated that only sentences in which the two meanings were represented by two distinct bracketings were disambiguated by durational cues. [ABSTRACT FROM AUTHOR]

Two paragraphs of computer-generated speech, with the same text, one made at Bell Laboratories by concatenating words spoken in isolation and the other at Haskins Laboratories using a phonetic dictionary, were evaluated with respect to the ease with which listeners could comprehend and retain information about the paragraph. A paragraph with the same text was recorded by a human speaker and used as a reference for comparison. Different groups of subjects listened to one of these three versions, hearing it one, two, or three times, and then answered questions concerning the content of the paragraph. The results indicate that for all three types of speech, comprehension measured in this way increased as a function of the number of times the paragraph was heard. The data indicated that there is a significant comprehensible difference between the two synthetic speech types. Scores for the speech generated by concatenating words spoken in isolation were closer to those for natural speech than were the scores obtained for speech synthesized from a phonetic dictionary. [ABSTRACT FROM AUTHOR]

Simple English sentences were generated by concatenating words spoken in isolation with minimum modifications of the synthesis parameters. The sentence forms were: Article-(adjective)-subject-verb-article-(adjective)-object. The two adjectives may or may not be present. A dictionary of 266 words specified by formant data with the duration of the words adjusted by linear interpolation to a reasonable duration for connected speech was stored. Pitch rules for the different sentences were derived from an average sentence pitch contour. Words were concatenated with no further duration adjustment, and formant smoothing was applied only at voiced adjacencies for one set of data of both the preceding and the following words; the pitch, however, was replaced by the rule pitch. The rule-generated sentences were compared with sentences obtained from formant analysis-synthesis of naturally spoken sentences, rather than with the naturally spoken sentences themselves, because formant analysis-synthesis introduces a characteristic nonnatural quality that is not a consequence of the rule-prosody. These comparisons demonstrated that the quality of the rule concatenated sentences is comparable to the quality of the resynthesized speech. [ABSTRACT FROM AUTHOR]