Concatenation Points Research Articles

Text Segmentation Via Processes that Count the Number of Different Words Forward and Backward

ABSTRACT The paper is developing a new statistical approach to automatic partitioning of texts into parts belonging to different authors. It is based on the analysis of processes that counts the number of different words forward and backward. The theoretical study of the processes is based on the assumptions of an elementary probability model with a change point. We prove consistence of our statistical estimate of the point of concatenation in the case when the concatenated texts have different Zipf exponents. This method is being tested on the Brown corpus and also on newspaper texts in different languages. Testing shows a good estimate of the concatenation point. This method can be used in parallel with other text segmentation methods.

A new fuzzy unit selection cost function optimized by relaxed gradient descent algorithm

In data-driven corpus-based text-to-speech synthesis systems, the main issue is to select the most natural-sounding sequence of acoustic units without unnatural acoustic transitions, and to minimize all acoustic mismatches at the concatenation points. Unit selection algorithms incorporating unit selection cost functions have been known to synthesize speech close to natural quality. However, these algorithms operate over large acoustic inventories with huge number of acoustic units in a broad spectrum of linguistic, prosodic and acoustic contexts, and with a huge number of concatenation possibilities. Moreover, the shape of the unit selection cost function, which evaluates the cost of concatenating two subsequent acoustic units, is modelled manually in a time-consuming and laborious iterative process, which is based on subjective evaluation. Since this process must be repeated for any new acoustic inventory, or even after changes in a given acoustic inventory, we propose instead a new fuzzy unit selection cost function. We further propose to optimize fully automatically the shape of the fuzzy unit selection cost function to the given acoustic inventory’s context by using a relaxed gradient descent algorithm, where the subjective tests are replaced by a novel objective measure needed to evaluate unit selection cost function performance. Furthermore, the proposed approach is fully interpretable and also highlights insights into which parts of the fuzzy unit selection cost function’s shape could be further improved. The experiments show that the optimized fuzzy unit selection cost function significantly outperforms the baseline fuzzy unit selection cost function. Moreover, the results prove that the unit selection optimization algorithm is capable of finding the optimal shape of the fuzzy unit selection cost function, even when optimized over a small subset of sentences.

Unit Selection Speech Synthesis Using Frame-Sized Speech Segments and Neural Network Based Acoustic Models

This paper proposes to select frame-sized speech segments for waveform concatenation speech synthesis using neural network based acoustic models. First, a deep neural network (DNN) based frame selection method is presented. In this method, three DNNs are adopted to calculate target costs and concatenation costs respectively for selecting candidate frames of 5ms length. One DNN is built in the same way as the DNN-based statistical parametric speech synthesis, which predicts target acoustic features given linguistic context inputs. The distance between the acoustic features of a candidate unit and the predicted ones for a target unit is calculated as the target cost. The other two DNNs are constructed to predict the acoustic features at current frame using its context features and the acoustic features of preceding frames. At synthesis time, these two DNNs are employed to calculate the concatenation cost for each candidate frame given its preceding ones. Furthermore, recurrent neural networks (RNNs) with long short-term memory (LSTM) cells are adopted to replace DNNs for acoustic modeling in order to make better use of the sequential information. A strategy of using multi-frame instead of single frame as the basic unit for selection is also presented to reduce the concatenation points within synthetic speech. Experimental results show that our proposed method can achieve better naturalness than the hidden Markov model (HMM)-based frame selection method and the HMM-based parametric speech synthesis method.

Chunk concatenation evolves with practice and sleep-related enhancement consolidation in a complex arm movement sequence.

This paper addresses the notion of chunk concatenation being associated with sleep-related enhancement consolidation of motor sequence memory, thereby essentially contributing to improvements in sequence execution speed. To this end, element movement times of a multi-joint arm movement sequence incorporated in a recent study by Malangré et al. (2014) were reanalyzed. As sequence elements differed with respect to movement distance, element movement times had to be purged from differences solely due to varying trajectory lengths. This was done by dividing each element movement time per subject and trial block by the respective “reference movement time” collected from subjects who had extensively practiced each sequence element in isolation. Any differences in these “relative element movement times” were supposed to reflect element-specific “production costs” imposed solely by the sequence context. Across all subjects non-idiosyncratic, lasting sequence segmentation was shown, and four possible concatenation points (i.e. transition points between successive chunks) within the original arm movement sequence were identified. Based on theoretical suppositions derived from previous work with the discrete sequence production task and the dual processor model (Abrahamse et al., 2013), significantly larger improvements in transition speed occurring at these four concatenation points as compared to the five fastest transition positions within the sequence (associated with mere element execution) were assumed to indicate increased chunk concatenation. As a result, chunk concatenation was shown to proceed during acquisition with physical practice, and, most importantly, to significantly progress some more during retention following a night of sleep, but not during a waking interval.

Digital Optoelectrical Pulse Method for Vernier-Type Rotary Encoders

This paper proposes the digital optoelectrical pulse (DOP) method for vernier-type rotary encoders in high-resolution applications. A 25-b encoder using the DOP method is demonstrated. In rotary encoders, the low-bit data of an angle measurement are easily generated using a periodic signal. The high-bit data of an absolute angle measurement can be obtained using various methods such as binary code, gray code, pseudorandom code, $M$ code, or the vernier method. The vernier method has a significant advantage over the above methods for small-sized rotary encoders because it requires fewer grating tracks than do other methods. The vernier method uses grating tracks to generate periodic signals rather than digitally coded signals. In the vernier method, periodic signals with slightly different frequencies are used. The high-bit data of an absolute angle measurement can be obtained using the difference of two periodic signals. Complete absolute angle data are obtained by concatenating the high-bit data and the low-bit data. However, the high-bit data and the low-bit data are, in general, not synchronized. As a result, significant intermediate-bit errors occur at the concatenation point. The DOP method is developed in this paper to mitigate these intermediate-bit error problems. A system on a chip (SoC) that uses the proposed method is designed and fabricated using a 0.18- $\mu \text{m}$ CMOS image sensor process. A 25-b absolute encoder is constructed using an SoC that provides absolute angle data without intermediate-bit errors.

The stuff that motor chunks are made of: Spatial instead of motor representations?

In order to determine how participants represent practiced, discrete keying sequences in the discrete sequence production task, we had 24 participants practice two six-key sequences on the basis of two pre-learned six-digit numbers. These sequences were carried out by fingers of the left (L) and right (R) hand with between-hand transitions always occurring between the second and third, and the fifth and sixth responses. This yielded the so-called LLRRRL and RRLLLR sequences. Early and late in practice, the keypad used for the right hand was briefly relocated from the front of the participants to 90° at their right side. The results indicate that after 600 practice trials, executing a keying sequence relies heavily on a spatial cross-hand representation in a trunk- or head-based reference frame that after about only 15 trials is fully adjusted to the changed hand location. The hand location effect was not found with the last sequence element. This is attributed to the application of explicit knowledge. The between-hand transitions appeared to induce initial segmentation in some of the participants, but this did not consolidate into a concatenation point of successive motor chunks.

Removal of Spectral Discontinuity in Concatenated Speech Waveform

synthesis systems which involve concatenation of recorded speech units are currently very popular. These systems are known for producing high quality, natural- sounding speech as they generate speech by joining together waveforms of different speech units. This method of speech generation is quite practical. However the speech units that are being concatenated may have different spectra on either side of the concatenation points. Such mismatches are spectral in nature and give rise to spectral discontinuity in concatenated speech waveforms. The presence of such discontinuities can be very distracting to the listener and degrade the overall quality of output speech. This paper proposes a speech signal processing technique that deals with the problem of spectral discontinuity in the context of concatenated waveform synthesis. It involves the post-processing of the synthesized speech waveform in time domain. This technique is implemented on different single channel Punjabi wave audio files which were created by concatenating different Punjabi syllables. A listening test was conducted to evaluate the proposed technique, and it was observed that the spectral discontinuity is reduced to a large extent and the output speech sounds more natural with the reduction of audible noise.

A Dynamic Cost Weighting Framework for Unit Selection Text–to–Speech Synthesis

Unit selection text-to-speech synthesis relies on multiple cost criteria, each encapsulating a different aspect of acoustic and prosodic context at any given concatenation point. Constraints are normally invoked on diverse characteristics such as inter-unit discontinuity, overall pitch contour, local duration profile, etc., leading to costs often too heterogeneous for a direct quantitative comparison. In order to rank available candidate units, this complexity must be reduced to a single number, and the relative importance of each information stream becomes highly critical. Yet this influence is typically determined in an empirical manner (e.g., based on a limited amount of synthesized data), yielding global weights that are thus applied to broad classes of concatenations indiscriminately. This paper proposes an alternative approach, <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">dynamic cost weighting</i> , based on a data-driven framework separately optimized for each concatenation considered. Specifically, the cost distribution in every stream is dynamically leveraged on a per concatenation basis to locally shift weight towards those characteristics that offer a high discrimination between candidate units, and away from those characteristics that are intrinsically less discriminative. An illustrative case study demonstrates the potential benefits of this solution, and listening evidence suggests that it does indeed entail higher perceived TTS quality.

Implementation of Phonetic Context Variable Length Unit Selection Module for Malay Text to Speech

Problem statement: The main problem with current Malay Text-To-Speech (MTTS) synthesis system is the poor quality of the generated speech sound due to the inability of traditional TTS system to provide multiple choices of unit for generating more accurate synthesized speech. Approach: This study proposes a phonetic context variable length unit selection MTTS system that is capable of providing more natural and accurate unit selection for synthesized speech. It implemented a phonetic context algorithm for unit selection for MTTS. The unit selection method (without phonetic context) may encounter the problem of selecting the speech unit from different sources and affect the quality of concatenation. This study proposes the design of speech corpus and unit selection method according to phonetic context so that it can select a string of continuous phoneme from same source instead of individual phoneme from different sources. This can further reduce the concatenation point and increase the quality of concatenation. The speech corpus was transcribed according to phonetic context to preserve the phonetic information. This method utilizes word base concatenation method. Firstly it will search through the speech corpus for the target word, if the target is found; it will be used for concatenation. If the word does not exist, then it will construct the words from phoneme sequence. Results: This system had been tested with 40 participants in Mean Opinion Score (MOS) listening test with the average rates for naturalness, pronunciation and intelligibility are 3.9, 4.1 and 3.9. Conclusion/Recommendation: Through this study, a very first version of Corpus-based MTTS has been designed; it has improved the naturalness, pronunciation and intelligibility of synthetic speech. But it still has some lacking that need to be perfected such as the prosody module to support the phrasing analysis and intonation of input text to match with the waveform modifier.

Applying the harmonic plus noise model in concatenative speech synthesis

This paper describes the application of the harmonic plus noise model (HNM) for concatenative text-to-speech (TTS) synthesis. In the context of HNM, speech signals are represented as a time-varying harmonic component plus a modulated noise component. The decomposition of a speech signal into these two components allows for more natural-sounding modifications of the signal (e.g., by using different and better adapted schemes to modify each component). The parametric representation of speech using HNM provides a straightforward way of smoothing discontinuities of acoustic units around concatenation points. Formal listening tests have shown that HNM provides high-quality speech synthesis while outperforming other models for synthesis (e.g., TD-PSOLA) in intelligibility, naturalness, and pleasantness.

Control of spectral dynamics in concatenative speech synthesis

Current speech synthesis methods based on the concatenation of waveform units can produce highly intelligible speech capturing the identity of a particular speaker. However, the quality of concatenated speech often suffers from discontinuities between the acoustic units, due to contextual differences and variations in speaking style across the database. In this paper, we present methods to spectrally modify speech units in a concatenative synthesizer to correspond more closely to the acoustic transitions observed in natural speech. First, a technique called "unit fusion" is proposed to reduce spectral mismatch between units. In addition to concatenation units, a second, independent tier of units is selected that defines the desired spectral dynamics at concatenation points. Both unit tiers are "fused" to obtain natural transitions throughout the synthesized utterance. The unit fusion method is further extended to control the perceived degree of articulation of concatenated units. A signal processing technique based on sinusoidal modeling is also presented that enables high-quality resynthesis of units with a modified spectral shape.

A Japanese TTS system based on multiform units and a speech modification algorithm with harmonics reconstruction

This paper proposes a new text-to-speech (TTS) system that utilizes large numbers of speech segments to produce very natural and intelligible synthetic speech. There are two innovations; new multiform synthesis units and a new speech modification algorithm based on a vocoder that offers harmonics reconstruction. The multiform units make it possible to reduce acoustic discontinuities at concatenation points and unnatural sound by preparing synthesis units with various lengths and various F/sub 0/ contours. The new speech modification algorithm, on the other hand, improves the quality of prosody modified speech. This algorithm is extremely effective in synthesizing speech whose prosodic parameters are quite different from those of synthesis units. Listening tests confirm that the new synthesis units yield speech with high intelligibility and naturalness, and that the new speech modification algorithm is superior to all other conventional vocoders and waveform domain algorithms including TD-PSOLA, especially when modifying the F/sub 0/ frequency upward.

Flexible post-lexical processing for speech synthesis from a large unit selection database

Online unit selection from large speech databases provides an opportunity to essentially play back words, phrases, and even sentences which were included in a recorded corpus. This capability can be extremely useful for limited domains, e.g., application prompts. Without switching voices, such a synthesizer could integrate high-quality synthesis with near-perfect recorded material. However, traditional post-lexical processing (PLP) considers only the phoneme specifications and not the sequences which actually exist in the target database. Phonemes supplied by the dictionary are typically rewritten into a single sequence of phones with reduced vowels, flapped t’s, etc. Given the enormous variability of human speech, any single sequence is unlikely to match an entire phrase or prompt as spoken and labeled. This paper addresses the use of flexible PLP, allowing multiple transcription possibilities which are essentially equivalent, at least for the speaker in question. By building the equivalences from the specific dictionary used by the synthesizer and the detailed phonetic labeling of a specific voice database, longer regions of the database can be selected, reducing the number of concatenation points in ordinary synthesis and increasing the odds of selecting complete recorded phrases.

The AT&amp;T Next-Gen TTS System

The new AT&amp;T TTS system for general U.S. English text is based on best-choice components picked from the AT&amp;T Flextalk TTS, the Festival System from the University of Edinburgh, and ATR’s CHATR system. From Flextalk, it employs text normalization, letter-to-sound, and (optionally) baseline prosody generation. Festival provides general software-engineering infrastructure (modularity) for easy experimentation and competitive evaluation of different algorithms or modules. Finally, CHATR’s unit selection was modified to guarantee the intelligibility of a good n-phone (n=2 would be diphone) synthesizer while improving significantly on perceived naturalness relative to Flextalk. Each decision made during the research and development phase of this system was based on formal subjective evaluations. For example, the best voice found in a test that compared TTS systems built from several speakers gave a 0.3-point head start (on a 5-point rating scale) in quality over the mean of all speakers. Similarly, using our Harmonic-plus-Noise speech representation gave us a 0.25-point advantage over standard TD-PSOLA. Finally, not performing prosodic modifications (other than some smoothing across concatenation points) on the units but using the system-generated prosody as a target in unit selection, 0.4 points were gained on overall quality. In conclusion, the new system combines the best of rule-based and data-driven worlds in TTS technology to deliver on the long-standing promise of truly natural-sounding synthesis.

Périodes locales et propagation de périodes dans un mot

A word of length n over an alphabet A is a sequence a 1 … a n of letters of A. It is convenient to consider a “long enough” word over A as an infinite right word, that is an infinite sequence a 1 … a i … of elements of A. An integer λ is a period of the word in the interval [ j … k] if we have a i = a i+ λ for those indices i and i + λ in the considered interval. The period of a word designates its smallest period over its whole length. A point p of a word is the cut ( a 1 … a p , a p+1 …). A non-negative integer λ is a local period at point p iff λ is a period in the interval [ p − λ + 1 … p + λ]. According to the critical point's theorem [1,2], the period of a “long enough (or not)” word is the maximum of the minimal local periods taken in each point of this word. M.P. Schützenberger, who was at the origin of the research work on the relations between local periods and periods of a word obtained by concatenation of periodical words, and our ability to characterize its period from the observation of the local period at the concatenation points only. This is the formulation of the unpublished answer we offered him that we suggest here.

A formant frequency modification algorithm dealing with the pole interaction

AbstractWhen a formant is modified based on the pole of the AR model, a problem arises due to the interaction between the pole to be modified and other poles such as the apparent elimination of the formant, if the frequency of the considered pole simply is modified. to solve this problem, the method proposed in this paper uses the spectral intensity at the pole frequency as a new feature parameter and the spectral intensity at the pole after the frequency modification is corrected to the desired value by iterative processings.To examine the effectiveness of the proposed method, an experiment was conducted in which a phoneme in a word is replaced into the same phoneme in a different environment, and then the formant frequency is modified by the proposed method in order to restore the original speech. the speeches before and after the modification are compared in terms of the cepstral distance (CD) from the original speech, and it is verified that the proposed method is effective in the modification of the formant. an attempt is made to solve the formant discontinuity problem at the concatenation point in waveform based speech synthesis‐by‐rule and to modify the formant frequency accompanying the modification of the fundamental frequency using the proposed method. It was verified as a result of experiment that the speech after formant modification is better than the speech before modification, indicating that the proposed method is useful in improving the quality of the synthesized speech.

An Algorithmic Interpretation of the GKS TEXT Primitive

The specification of text output is one of the most complex parts of the GKS standard [1]. The text “geometry” is determined by a wide variety of factors: some, such as character height, are specified directly by the user; others, such as the aspect ratio of a character in a given font, are characteristic of a particular implementation. The GKS document gives a detailed, but informal, description of the way in which these various factors combine to determine the text size and position. The purpose of this paper is to put forward a more formal description of the GKS text geometry: an algorithm is presented that calculates the size and positions of individual character bodies for any text string, under any setting of text attributes, together with the text extent rectangle and concatenation point; a realization of the algorithm as a FORTRAN 77 subroutine is also given. It is hoped that the algorithm will help implementos to interpret the GKS standard more easily.