Phoneme Duration Research Articles

The effect of variations in method of elicitation on maximum sustained phoneme duration

This investigation evaluated the effect of method of elicitation on the maximum sustained durations of /s/, /z/, and /a/ in eighty female subjects. The effect of order of elicitation on the maximum sustained durations of /s/ and /z/, and the effects of method and order of elicitation on the s/z ratio, were analyzed as well. Comparisons of three methods of measuring maximum phoneme duration (MPD) were performed. This study found method of elicitation of MPD to have an effect on the maximum sustained durations of /s/, /z/, and /a/ but not on the s/z ratio. The three methods of measuring MPD were found to correlate highly with one another. Results of this study point to the need for standardized procedures for elicitation of MPD.

Real Ear Compression Ratios

An investigation was conducted to determine the effects of vent diameter, stimulus characteristics, and adaptive release time on the real ear compression ratio (CR). Real ear input/output (I/O) functions generated in response to syllables and speech-shaped noise revealed that the real ear CR is often less than the nominal value measured in the coupler, but only for CRs greater than 2:1. Results were similar across all vents except IROS, a condition that resulted in much larger errors than smaller vents. Pooled across vents, CRs of individual phonemes were found to be associated with phoneme duration, crest factor, and the interval of time between phonemes. Results were also influenced by an adaptive release time, which may have affected the compression of vowels and consonants differently.

The contribution of intonation, segmental durations, and spectral features to the perception of a spontaneous and a read speaking style

The influence of pitch contour, segmental durations, and spectral features on the perception of two speaking styles was studied. For this purpose two male speakers each spoke “spontaneously” to an interviewer and afterwards read out their own literally transcribed spontaneous text. Pairs of identical spontaneous and read utterances were selected that were fluently spoken in both speaking styles (no false starts, hesitations, etc.). Five test conditions were constructed in which the utterances had: (1) no manipulations; (2) phoneme durations from the opposite speaking style; (3) the pitch contour from the opposite speaking style; (4) a monotonous pitch contour; (5) the original spectral features combined with the prosodic features of the opposite speaking style. The stimuli were presented to 32 subjects in a listening experiment. Their task was to classify each utterance as either “spontaneous” or “read out” speech. All manipulations of the test utterances had a significant effect on the classification of the speaking style. We also analysed the original utterances with respect to several acoustic measures for intonation, duration, jitter and shimmer, and spectral vowel quality. Overall, read speech compared to spontaneous speech had: a lower articulation rate, more F 0 variation, more F 0 declination, less shimmer, and less vowel reduction. However, none of these acoustic features by itself can clearly discriminate between the two speaking styles. Above all it became clear that the performance of the speakers and the listeners varied enormously.

Helium speech database and some new aspect of diver’s speech

Speech is distorted in pressurized helium-air. Many aspects, except for upward shifts of formant and pitch in the helium speech, however, have not been known yet. A helium speech database system including more than 5 000 words and messages of divers was developed. The vibration of the cheek, the neck, and the nose were also recorded for some utterances. They were taken at the simulation dives up to the depth of 300 m at JAMSTEC since 1982. The helium speech and the normal speech were compared in terms of (1) duration of phonemes, and (2) pitch frequencies and formant trajectories in two different hearing/speaking conditions: the first condition is a case when a diver spoke as he was listening to his own unscrambled speech, and in the other condition the diver was not hearing his unscrambled speech. As a result, the shorter duration of word-initial consonants, higher pitch by psychological stress, and rough movement of formant trajectories were observed in addition to the well-known formant and pitch frequency shifts. As a special property of helium speech, it is newly discovered that there exists a quasiformant below the normal first formant, and that it may be caused by the sound from the nostrils.

Phoneme recognition using reference patterns constructed with discriminative training and DP matching

A modified LVQ2 algorithm (MLVQ2) for phoneme recognition is proposed. The length of a phoneme reference pattern made by the algorithm was fixed. It showed high phoneme recognition performances for isolated spoken words. However, it did not show high performances for continuous speech because the phoneme reference patterns with fixed length could not deal with a big change in phoneme duration. In this paper, a new construction algorithm of the reference patterns using discriminative training based on the MLVQ2 and DP matching is proposed. The initial multiple reference patterns are optimized as follows: At first, the correspondence between a frame of the input sample and a frame of the nearest reference pattern in the incorrect phoneme class is calculated by using the DP matching. Then, the corresponding frame vector of the reference pattern in the incorrect phoneme class is moved away from the frame vector of the input sample based on the MLVQ2. On the other hand, the corresponding frame vector of the reference pattern of the correct phoneme class is moved nearer to the frame vector of the input sample. The experimental result using the ASJ 503 sentences speech database showed a 81.3% phoneme recognition rate, increasing 1.7% compared to the previous method.

Training section tolerant HMM in concatenated training

Concatenated training of phoneme HMMs can use speech data without hand labels for training HMMs, but it has a tendency to decrease the recognition rate because of an improper training section. The results for speaker-independent isolated word recognition experiments showed that non left–right HMM is tolerant of the quality of the training section in concatenated training than the conventional left-to-right HMM. Non left–right HMM has a structure where state transitions within a phoneme are ergodic and state transitions between successive two phonemes are left-to-right. Non left-right HMM and the conventional left-to-right HMM show much the same performance as long as the training section is given properly by hand labels. Since the training section contains undesirable data, the recognition rate for both HMMs decreases, but the decrement of the recognition rate for non left-right HMM is smaller than the decrement for the conventional left-to-right HMM. Similar tendency was shown when the training section was determined by a phoneme duration model independently from the hand labels.

The use of prosody in automatic speech recognition

The acoustic correlates of prosody, energy, duration, and fundamental frequency are easy to extract from the speech signal. Yet, prosody has virtually been ignored in large vocabulary speech recognition. Many systems use energy and its time derivatives in addition to spectral information, and some systems use phoneme duration models. However, those cannot truly be referred to as prosodic models. There are many reasons why it has been difficult to successfully use models of prosody in speech recognition: (a) Speech recognition models are based on phoneme-sized acoustic units, while prosodic information spans several levels, ranging from subphonetic, to phrase level; (b) speech recognition assumes quasistationarity and primarily deals in terms of absolute values, while prosodic models are susceptible to such effects as speaking rate and the speakers F0 range, requiring analysis in terms of relative values, which can cause problems in one-pass real-time systems; (c) the benefits of prosodic modeling in speech recognition are moderate (for English), and the computational cost can be huge. However, there are potential solutions for most of these problems, and as the speech recognition problem gradually becomes a speech understanding problem, more aspects of prosodic modeling will find their way into the speech recognition and understanding systems.

On the use of a hybrid harmonic/stochastic model for TTS synthesis-by-concatenation

In this paper, we address the possibilities offered by hybrid harmonic/stochastic (H/S) models in the context of wide-band text-to-speech synthesis based on segment concatenation. After a brief recall of the hypotheses underlying such models and a comprehensive review of a well-known analysis algorithm, namely the one provided by the multi-band excited (MBE) analysis framework, we study how H/S models allow to modify the prosody of segments and how segment concatenation can be organized, in the purpose of minimizing mismatches at the border of segments. In this context, we introduce an original concatenation algorithm which takes advantage of some analysis errors. Speech synthesis algorithms are then described, including an original synthesis technique based on judiciously prepared IFFTs, and the final segmental quality 1 1 Segmental quality, in TTS synthesis, is the ability of a synthesizer to produce natural-sounding speech sounds. It is opposed to supra-segmental quality, which relates to the adequacy of the pitch and phoneme durations computed by the language processing block of the TTS system to the sentence to be synthesized. is detailed. More particularly, we examine the differences in the quality obtained when using the model in a narrow-band speech coding context and in a wide-band, concatenation based synthesis context. We study three possible causes for these differences: the choice of an analysis criterion, the inadequacy of the model due to pitch variatons, and the effect of coarticulation on phases.

Perceptual normalization for speaking rate: effects of temporal distance.

A series of studies was undertaken to examine how rate normalization in speech perception would be influenced by the similarity, duration, and phonotactics of phonemes that were adjacent or distal from the initial, target phoneme. The duration of the adjacent (following) phoneme always had an effect on perception of the initial target. Neither phonotactics nor acoustic similarity seemed to have any influence on this rate normalization effect. However, effects of the duration of the nonadjacent (distal) phoneme were only found when that phoneme was temporally close to the target. These results suggest that there is a temporal window over which rate normalization occurs. In most cases, only the adjacent phoneme or adjacent two phonemes will fall within this window and thus influence perception of a phoneme distinction.

Age-related differences in speech variability among women

The purpose of this study was to examine the differences in variability of several measures of the speech produced by two groups of women, one aged 20–35 years and one aged 75 years and over. The subjects read the first paragraph of the “Rainbow Passage,” sustained vowels at three loudness levels, and repeated a carrier phrase in which a series of 18 syllables were embedded. Data were gathered on intraoral pressure ( PIO), vocal intensity, speaking fundamental frequency (SFF), voice onset time (VOT), and phoneme duration. The F-max test and ANOVA statistics revealed that the older women exhibited greater within-subject variability than the younger women for PIO, SFF, VOT, and consonant duration. These factors required more precise temporal, respiratory, phonatory, or articulatory adjustments than did the vocal intensity tasks or vowel duration. The variability differences found in this study may reflect individual patterns of acoustic and physiologic change in the speech produced by older women.

Analysis of prominence in spoken Japanese sentences and application to text-to-speech synthesis

This paper focuses on the partial emphasis or “prominence” of parts of Japanese sentences. Four sets of 43 read sentences uttered by two speakers including various types of prominence (172 sentences in total) are analyzed. This analysis shows that in 88% of the sentences prominence is produced by enhancing F 0 and increasing power. No examples of lengthening of phoneme duration are observed in the emphasized parts of the sentences except for some special cases. One exception is lengthening accompanied by pause insertion as a mark of prominence, and another slowing total speech rate. The prosodic features of read natural speech are then used to develop rules for changing a reference sentence to produce prominence for rule-based speech synthesis. Listening test results using 10 subjects do not show any significant difference in expressibility between prominence synthesized by rule (rate of correct expression: 76.9%) and prominence in natural speech (79.9%) at the 5% level. To further improve prominence expressibility, listening tests for 10 subjects are used to clarify the conditions under which prominence expressibility becomes optimal. These tests show that the prosodic control parameters increase the expressibility of prominence by about 20%. Finally, prosodic features of spontaneous conversational speech are analyzed and compared with those of read sentence speech. Speech-rate reduction in parts where prominence is placed is more conspicuous in spontaneous conversational speech.

Method of and apparatus for speech recognition wherein decisions are made based on phonemes

Linear prediction coefficients of a speech signal including unknown words are derived for each of successive periodic frame intervals. For every frame over the duration of an individual phoneme of the speech signal, the degree of similarity of stored coefficients of known words and derived coefficients of the unknown words are calculated so that at the end of the individual phonemes, the degree of similarity is calculated. Phoneme segmentation data are derived in response to the speech signal and combined with the calculated degree of similarity over the individual phoneme to derive phoneme strings of the speech signal. The derived and stored phoneme strings are compared to indicate the words stored in a word dictionary having the greatest similarity with the derived phoneme strings.

Control of phoneme duration based on the movement of speech organs.

Control of phoneme duration based on the movement of speech organs.

Use of minimum duration and energy contour for phonemes to improve large vocabulary isolated-word recognition

Many acoustic misrecognitions in our 86 000-word speaker-trained isolated-word recognizer are due to phonemic hidden Markov models (phoneme models) mapping to short segments of speech. When we force these models to map to longer segments corresponding to the observed minimum durations for the phonemes, then the likelihood of the incorrect phoneme sequences drops dramatically. This drop in the likelihood of the incorrect words results in significant reduction in the acoustic recognition 1 1 We use the term acoustic recognition error rate to mean the recognition error rate when every word in the vocabulary is considered a priori equally likely. error rate. Even in cases where acoustic recognition performance is unchanged, the likelihood of the correct word choice improves relative to the incorrect word choices, resulting in significant reduction in recognition error rate with the language model. On nine speakers, the error rate for acoustic recognition reduces from 18·6 to 17·3%, while the error rate with the language model reduces from 9·2 to 7·2%. We have also improved the phoneme models by correcting the segmentation of the phonemes in the training set. During training, the boundaries between phonemes are not marked accurately. We use energy to correct these boundaries. Application of an energy threshold improves the segment boundaries between stops and sonorants (vowels, liquids and glides), between fricatives and sonorants, between affricates and sonorants and between breath noise and sonorants. Training the phoneme models with these segmented phonemes results in models which increase recognition accuracy significantly. On two speakers, the error rate for acoustic recognition reduces from 26·5 to 23·1%, while the error rate with the language model reduces from 11·3 to 8·8%. This reduction in error rate is in addition to the error rate reductions obtained by imposing minimum duration constraints. The overall reduction in errors for these two speakers using minimum durations and energy thresholds is from 27·3 to 23·1% for acoustic recognition, and from 14·3 to 8·8% with the language model.

A PC-based very large vocabulary isolated word speech recognition system

The Speech Recognition Group at Olivetti has designed an isolated word recognition system for Italian which can accept a very large vocabulary in excess of 60,000 words and recognize words in real time. The system fits into a 386-based personal computer and runs under MS-DOS. Among its most appealing features are the very quick and application-independent procedure for training the system to the user's voice and the easy way in which application dictionaries can be created. What makes short training possible is the fact that the system extracts target spectral distributions of phonemes from the training words and combines this speaker-dependent information with speaker-independent information concerning other aspects of the speech signal such as energy profile, phoneme durations, and phoneme similarity. The system accepts natural language without any restriction and can be used to dictate documents of any kind, provided the dictionary of the application domain is specified. Dictation, however, is not the only application we envisage. In this paper we describe our approach to speech recognition and provide some information on the actual implementation and performance of our system.

A Japanese text‐to‐speech conversion system using pitch‐controlled residual wave excitation

A Japanese text‐to‐speech conversion system was developed that could generate fluent speech by concatenating CV and VC units. A new synthesis method was proposed, where residual waves were used as excitation signals for an LPC synthesis filter in all portions of each unit. LPC filter coefficients were calculated by approximating spectral envelopes extracted by the improved cepstral analysis method, which was less affected by pitch frequency than the conventional LPC method. Therefore, the synthetic speech generated by the proposed method had extremely high quality, even when pitch frequencies of the residual waves were widely changed. Moreover, in order to realize natural rhythms, phoneme duration was determined by new rules. The rules were created by taking into account the various contexts (such as preceding and following phonemes and the position of a phoneme in a phrase), and by statistically analyzing a large speech database. A comprehension test using 1000 phonetic balanced words showed that the proposed system achieved 97.4% of high accuracy rate.

Duration control methods for HMM phoneme recognition

Two kinds of duration control for HMM (hidden Markov model) phoneme recognition are proposed: phoneme duration control for an HMM phone model and an event duration control for an HMM state. The phoneme duration control is carried out by combining an HMM output probability with a phoneme duration probability. The phoneme duration probability is calculated using a phoneme duration histogram obtained from training samples. Phoneme duration control is effective in discriminating phonemes with different durations such as /n/ and /N/. Event duration control is realized as a state duration penalty calculated from an HMM state duration probability of training samples. Event duration control is effective in discriminating phonemes with different event structures such as /s/ and /ts/. Recognition experiments are carried out using Japanese phonemes extracted from an isolated word database uttered by one male speaker. The phoneme recognition rate is improved from 84.8%–90.0% using these duration control techniques.

A method for selecting an optimum phoneme sequence using a posteriori probabilities of phonemes

A phoneme recognition system using phoneme discriminant filters was proposed, where segmentation and phoneme recognition were carried out simultaneously. However, the method had many phoneme insertion errors and often produced phoneme sequences that were inconsistent with Japanese phoneme connection rules. This paper proposes methods for reducing the number of phoneme insertion errors using inhibitory functions and for selecting an optimum phoneme sequence using dynamic programming based on the a posteriori probabilities of the phonemes. The definition of inhibitory functions is based on the Japanese phoneme connection rules and the phoneme duration information. This method was tested for the recognition of phonemes in a 212‐work vocabulary uttered by ten male speakers. The inhibitory functions reduced the total number of phoneme insertion errors to 1/5 and, especially, reduced the number of consonant insertion errors. The a posteriori probabilities of phonemes were computed based on distributions of the output of the phoneme discriminant filters.

Analysis and synthesis of the singing voice and mnsical applications

IRCAM (Institut de Recherche et de Coordination Acoustique/Musique) has been involved in sound analysis and synthesis for contemporary music creation for many years. Accurate analysis and high‐quality sound synthesis are thus extensively studied. Some of the methods that have been developed are described with particular attention to the singing voice and speech. A power spectral density estimate has been developed of the speech signal, which is simultaneously very precise and computed on very short segments of the signal in order to follow rapid transitions such as those found in plosives. Spectral envelopes are represented in a flexible way: They are coded in terms of frequency, amplitude, and bandwidth of pairs of conjugate poles (formants). An algorithm has been developed giving an harmonic/noisy decision in narrow frequency bands covering the entire spectrum. This technique has greatly improved the sound quality. Several synthesis methods have been developed and compared, including formant waveform synthesis (analogous to parallel formant synthesis), additive synthesis, and serial filter synthesis. Two types of synthesis rules have been studied. On the one hand, rules concerning phoneme coarticulation using formant coded diphones, sophisticated algorithm smoothing formant trajectories by taking into account three successive phonemes, durations of phonemes, speed of singing, speed of transitions, and relative timing of formant trajectories. On the other hand, a set of rules has been built with musical performance. Those rules calculate such features as vibrato excursion and rate, vocal effort, and formant locations according to fundamental frequency. Examples of synthesis by rule are given showing different types of voices and styles plus excerpts from concert pieces.

Delta streams for English synthesis

Although speech patterns are heavily influenced by the hierarchical linguistic structure of speech, synthesis by rule systems have generally been based on linear utterance representations. Delta is a new programing language that makes it easy to work with utterance representations containing multiple levels. Both higher level linguistic units, such as phrases, syllables, and phonemes, and lower level phonetic events, such as articulatory or formant targets and F0 trends, can be easily accommodated on separate interconnected “streams,” with each unit equally accessible to the rules. While Delta can be used to test most any synthesis model for any language, this paper will show how Delta can be used to test a particular model for English. This model uses, among others, CV, syllable, nucleus, phoneme, formant, and duration streams, with formant transitions represented as duration tokens that are in effect invisible in other streams. The paper will justify the selection of streams and the unique way of handling formant transitions, demonstrating in Delta notation how the model leads to particularly straightforward rules for predicting English phoneme durations, formant values, and aspiration patterns.