Good Speech Quality Research Articles

Iterative spectral subtraction method for millimeter-wave conducted speech enhancement

A non-air conducted speech detecting method has been developed in our laboratory by using millimeter wave radar technology. Because of the special attributes of the millimeter wave, this method may considerably extend the capabilities of traditional speech detecting methods. However, radar speech is substantially degraded by additive combined noises that include radar harmonic noise, electrocircuit noise, and ambient noise. This study, therefore, proposed an iterative spectral subtraction method which can be adaptively estimate noise spectrum at every iteration, and reduce the musical noise remained in the previous spectral subtraction process. Results from simulations as well as evaluations confirm that the proposed method satisfactorily reduces whole-frequency and musical noises and produces good speech quality.

Robust Speaker-Adaptive HMM-Based Text-to-Speech Synthesis

This paper describes a speaker-adaptive HMM-based speech synthesis system. The new system, called ldquoHTS-2007,rdquo employs speaker adaptation (CSMAPLR+MAP), feature-space adaptive training, mixed-gender modeling, and full-covariance modeling using CSMAPLR transforms, in addition to several other techniques that have proved effective in our previous systems. Subjective evaluation results show that the new system generates significantly better quality synthetic speech than speaker-dependent approaches with realistic amounts of speech data, and that it bears comparison with speaker-dependent approaches even when large amounts of speech data are available. In addition, a comparison study with several speech synthesis techniques shows the new system is very robust: It is able to build voices from less-than-ideal speech data and synthesize good-quality speech even for out-of-domain sentences.

REAL-TIME SPEECH COMPRESSION BY USING CODE EXCITED LINEAR PREDICTION ALGORITHM

A lot of effort has been spent over the last few years in the development of digital speech coding methods and their subsequent standardization. Algorithms have evolved which provide good quality speech at sub 8 kbps bit rates although at a much computational expense. Speech compression is proposed based on code excited linear prediction algorithm and implementation in DSP algorithm. Algorithm based on three-stage technique which involves simulate, evaluate, debug and implementation in G.723 low delay code excited linear prediction (LD-CELP)[4] algorithm. First stage, the algorithm is evaluated via simulation to determine whether it meets the design criterion. Then, it is implemented in real-time based an object oriented approach. After the algorithm is thoroughly tested, it is further refined to obtain tighter and faster coding. This technique can be applied to other real-time DSP algorithms. A simulation result shows that better speech quality is obtained. The techniques described in this paper are applicable to any other speech codec.

Wavelet packet entropy for millimeter wave conducted speech enhancement

A new nonacoustic sensor-based speech detecting method has been developed in our laboratory by using millimeter wave radar technology. Because of the special attributes of the millimeter wave, this method may considerably extend the capabilities of traditional speech detecting methods. However, radar speech is substantially degraded by additive combined noises that include radar harmonic noise, electrocircuit noise, and ambient noise. Therefore, this study proposes an improved spectrum subtraction algorithm that uses wavelet packet entropy analysis to detect voiced and unvoiced sections in an adaptive manner. Use of this algorithm leads to a significant reduction in the unnatural components of residual noise. Results from simulations as well as evaluations confirm that enhanced nonacoustic sensor speech is more pleasant to listeners than speech produced by traditional spectral subtraction algorithms. This suggests that the proposed method satisfactorily reduces whole-frequency and musical noises and produces good speech quality

A time-frequency blind separation method for underdetermined speech mixtures

The proposed Blind Source Separation method (BSS), based on sparse representations, fuses time-frequency analysis and the clustering approach to separate underdetermined speech mixtures in the anechoic case regardless of the number of sources. The method remedies the insufficiency of the Degenerate Unmixing Estimation Technique (DUET) which assumes the number of sources a priori. In the proposed algorithm, the Short-Time Fourier Transform (STFT) is used to obtain the sparse representations, a clustering method called Unsupervised Robust C-Prototypes (URCP) which can accurately identify multiple clusters regardless of the number of them is adopted to replace the histogram-based technique in DUET, and the binary time-frequency masks are constructed to separate the mixtures. Experimental results indicate that the proposed method results in a substantial increase in the average Signal-to-Interference Ratio (SIR), and maintains good speech quality in the separation results.

A New Buffer Algorithm for Speech Quality Improvement in VoIP Systems

Jitter buffer plays an important role in Voice over IP (VoIP) applications because it provides a key mechanism for achieving good speech quality to meet technical and commercial requirements. The main objective of this paper is to propose a new, simple-to-use jitter buffer algorithm as a front-end to conventional static or adaptive jitter buffer algorithms to provide improved performance, in terms of enhanced user-perceived speech quality and reduced end-to-end delay. Supported by signal processing features, the new algorithm, the so-called Play Late Algorithm, alters the playout delay inside a speech talkspurt without introducing unnecessary extra end-to-end delay. The results show that the new algorithm achieves the best performance under different network conditions when compared to conventional static and adaptive jitter buffer algorithms. The results reported here are based on live tests and emulated network conditions on real mobile phone prototypes. The mobile phone prototypes use AMR codec and support full IP/UDP/RTP stack with IPSec function in some of the tests. The method for perceived speech quality measurement is based on the ITU-T standard for speech quality evaluation (PESQ).

Speech enhancement based on undecimated wavelet packet-perceptual filterbanks and MMSE–STSA estimation in various noise environments

In this paper, we proposed a new speech enhancement system, which integrates a perceptual filterbank and minimum mean square error–short time spectral amplitude (MMSE–STSA) estimation, modified according to speech presence uncertainty. The perceptual filterbank was designed by adjusting undecimated wavelet packet decomposition (UWPD) tree, according to critical bands of psycho-acoustic model of human auditory system. The MMSE–STSA estimation (modified according to speech presence uncertainty) was used for estimation of speech in undecimated wavelet packet domain. The perceptual filterbank provides a good auditory representation (sufficient frequency resolution), good perceptual quality of speech and low computational load. The MMSE–STSA estimator is based on a priori SNR estimation. A priori SNR estimation, which is a key parameter in MMSE–STSA estimator, was performed by using “decision directed method.” The “decision directed method” provides a trade off between noise reduction and signal distortion when correctly tuned. The experiments were conducted for various noise types. The results of proposed method were compared with those of other popular methods, Wiener estimation and MMSE–log spectral amplitude (MMSE–LSA) estimation in frequency domain. To test the performance of the proposed speech enhancement system, three objective quality measurement tests (SNR, segSNR and Itakura–Saito distance (ISd)) were conducted for various noise types and SNRs. Experimental results and objective quality measurement test results proved the performance of proposed speech enhancement system. The proposed speech enhancement system provided sufficient noise reduction and good intelligibility and perceptual quality, without causing considerable signal distortion and musical background noise.

Nasalance in patients with maxillary defects – Reconstruction versus obturation

An important question remains on how to obtain good quality of speech for patients needing maxillectomy. Oral and nasal spaces must be separated either by surgical means or by using an obturator-prosthesis. An objective measure of oronasal closure is nasalance. Different rehabilitative strategies should be compared. Between 1990 and 2000, 88 patients underwent maxillectomy of which 28 (32%) were available for examination. Ten patients had obturators (group l) and in 18 patients the maxilla was biologically reconstructed with different techniques (group 2). Sound pressure of nasal and oral airways were assessed seperately using a computerized sampling system (NasalView) and standardized German texts. Nasalance was calculated and compared with an uncompromised sample of patients. There were no significant differences between group 1 and group 2 concerning nasalance. Furthermore, the achieved values of nasalance were similar to healthy individuals. Nasalance after maxillectomy can be normal after sufficient rehabilitation.

Późne foniatryczne wyniki leczenia chorych z rozszczepami podniebienia operowanych przez jednego chirurga

The aim of the study was the assessment of the late speech results the cleft palate patients. MATERIAL AND METHODIC. From January 1987 to December 1996 the first author performed 268 cleft palate surgeries. Of this number, 190 were complete unilateral and bilateral clefts, while 78 were operations of isolated soft palate. All the patients were consulted by a phoniatrist and a plastic surgeon. Among others, articulation, nasality, speech intelligibility were evaluated and nasometirc investigations were performed. The follow up examination involved 133 patients, 56 females and 77 males aged from 9 to 19 years (mean at the age of 13)--61 patients with unilateral cleft lip, alveolar process and palate (UCLP), 30 with bilateral clefts (BCLP) and 42 with isolated soft palate clefts (ICP). All the patients were operated on between 20 and 36 months of age (mean at 28 months). Two patients (1.5%) developed dehiscence of the wound on the palate and further two (1.5%) developed fistulae localized behind the alveolar process at a later time. All children with clefts were operated on by classical methods with the use of mucoperiosteal flaps for complete clefts according to Veauís method and soft palate clefts--Kilner-Wardill's method. Very good and good results of treatment were observed in 92% of cases. Poor outcome of treatment--excessive nasality or poor speech quality was observed in 11 patients (8%). A similar evaluation of the outcome of treatment was obtained on the basis of questionnaires sent to patients. It seems that good speech quality can be achieved despite relatively late palate repair using Veau and Kilner-Wardill methods.

Recording high quality speech during tagged cine‐MRI studies using a fiber optic microphone

To investigate the feasibility of obtaining high quality speech recordings during cine imaging of tongue movement using a fiber optic microphone. A Complementary Spatial Modulation of Magnetization (C-SPAMM) tagged cine sequence triggered by an electrocardiogram (ECG) simulator was used to image a volunteer while speaking the syllable pairs /a/-/u/, /i/-/u/, and the words "golly" and "Tamil" in sync with the imaging sequence. A noise-canceling, optical microphone was fastened approximately 1-2 inches above the mouth of the volunteer. The microphone was attached via optical fiber to a laptop computer, where the speech was sampled at 44.1 kHz. A reference recording of gradient activity with no speech was subtracted from target recordings. Good quality speech was discernible above the background gradient sound using the fiber optic microphone without reference subtraction. The audio waveform of gradient activity was extremely stable and reproducible. Subtraction of the reference gradient recording further reduced gradient noise by roughly 21 dB, resulting in exceptionally high quality speech waveforms. It is possible to obtain high quality speech recordings using an optical microphone even during exceptionally loud cine imaging sequences. This opens up the possibility of more elaborate MRI studies of speech including spectral analysis of the speech signal in all types of MRI.

Short low- and medium-frequency antenna performance

Lately, short antennas have attracted the attention of the broadcast and communication communities. This kind of antenna has been used since the 1920s. Top-loaded monopoles are the logical antennas to be used in order to get a low-profile antenna and performance according to the needs of the broadcaster and for communication. In this paper, top-loaded monopoles have been studied exhaustively using the transmission-line technique. Improved expressions for the antenna's radiation resistance have been obtained, taking into account the top-base current relationship and under different top-loading conditions. This idea, of using an equivalent transmission-line technique, has been used since the 1920s in order to obtain the antenna's input reactance. Using this old idea, the novel approach here permits obtaining the near- and far-field expressions from the current distribution on the antenna structure. A near-field calculation is used to determine the surface-current density on the ground plane. The power dissipation is calculated from the artificial and natural ground-plane surface-current densities, and the ground plane equivalent loss resistance is obtained. In all cases, as a first approximation, a half-wavelength ground-plane radius has been used, because this is the maximum distance covered by the ground surface current under the antenna, closing the antenna's electric circuit. Beyond this distance, the ground currents do not return to the antenna generator, and are taken into account in the surface-wave propagation calculations. The half-wavelength ground-plane surface is partially occupied by the metallic radial ground system, and the remainder consists of the natural soil. Artificial ground-plane behavior is paramount in obtaining the best performance for a short antenna. This kind of antenna could perform very close to a standard quarter-wave monopole, if it has optimum dimensions. For these reasons, a short antenna and the corresponding artificial ground plane have been analyzed, modifying the number of radials and their lengths, in order to achieve optimum performance, or to obtain maximum field strength with several soil conditions for the Earth's surface. A very simple and efficient antenna can be obtained, giving to the broadcast and communication communities a product that can fulfil the required performance of radiating a high-quality AM signal or digital transmission in the MF band, and good speech quality in the LF band.

A sinusoidal voice over packet coder tailored for the frame-erasure channel

A speech coder tailored especially for the frame-erasure channel-the sinusoidal voice over packet coder (SVOPC)-is proposed. Based on a classified approach, avoiding interframe coding techniques, and synthesizing its output from slowly varying parameters, the coder is inherently robust to packet loss. SVOPC is based on quasi-harmonic modeling of the linear prediction (LP) residual. Both the sinusoidal amplitudes and phases are explicitly encoded using new methods based on Gaussian mixture models. A wide-band (16-kHz sampling frequency) implementation of the coder provides synthesized speech of good subjective quality at around 20 kbps. SVOPC is evaluated by means of subjective listening tests, and compared to a reference system based on G.722.2 (the AMR wide-band codec). Under frame erasure conditions (5%-30% frame erasures generated according to a Gilbert model), SVOPC clearly outperforms G.722.2.

Matrix quantization and mixed excitation based linear predictive speech coding at very low bit rates

A matrix quantization scheme and a very low bit rate vocoder is developed to obtain good quality speech for low capacity communication links. The new matrix quantization method operates at bit rates between 400 and 800 bps and using a 25 ms linear predictive coding (LPC) analysis frame, spectral distortion about 1 dB is achieved at 800 bps. Techniques for improving the performance at very low bit rate vocoding include quantization of residual line spectral frequency (LSF) vectors, multistage matrix quantization, joint quantization of pitch and voiced/unvoiced/mixed decisions and a technique to obtain voiced/unvoiced/mixed decisions. In the new matrix quantization based mixed excitation (MQME) vocoder, the residual LSF vectors for two consecutive frames are obtained using autoregressive moving average (ARMA) prediction, then grouped into a superframe and jointly quantized. For other speech parameters, quantization is made in each frame. The residual LSF vector quantization yields bit rate reduction in the vocoder. For the MQME vocoder, listening tests have proven that an efficient and high quality coding has been achieved at a bit rate of 1200 bps. Test results are compared with the mixed excitation based 2400 bps MELP vocoder which is chosen as the new federal standard, and it is observed that the degradation in speech quality is tolerable and the performance is near the 2400 bps MELP vocoder particularly in quiet environments.

Fast SOLA-Based Time Scale Modification Using Envelope Matching

Time scale modification (TSM) of speech and audio signals is very useful in many applications such as MPEG-4 and fast/slow browsing of pre-recorded materials. Synchronized Overlap-and-Add (SOLA) is a time-domain TSM algorithm known to achieve good speech and audio quality. One problem of SOLA is that it requires a large amount of computation in the search of the best matching point between the analysis and synthesis frames. In this paper, we propose two algorithms, envelope-matching TSM (EM-TSM) and modified EM-TSM (MEM-TSM), to simplify the computation with negligible perceptual quality degradation. In EM-TSM, 1-bit sign information is used in the search to substitute the full-precision signal samples used in SOLA. Three additional computation reduction measures, namely simplified formulation, recursive computation and search-point reduction, are applied to achieve significant computation reduction. In MEM-TSM, we reduce the computation of EM-TSM further by introducing zero-crossing point reduction and predictive search skipping. We also improve the quality of the time-scaled signals by introducing multiple-candidate re-examination, and frame-size modification. Simulation results show that the proposed MEM-TSM can achieve computational reduction factors as large as 300 with very good perceptual quality of time-scaled speech and audio.

Coding Method of LSP Residual Signals Using Wavelets for Speech Synthesis

This paper presents a method to use wavelet analysis for speech coding and synthesis by rule. It is a coding system where LSP residual signal is transformed into wavelet coefficients. As wavelet analysis is implemented effectively by filter banks, our method is featured to require less computation than multipulse coding and others where complicated prediction procedures are essential. To achieve good quality speech at low bit rates, we verified to allocate the different bits onto the wavelet coefficients, with more bits in lower frequencies, and less in higher. The synthesized speech by Haar wavelet with 16.538kbits/s has nearly same perceptual quality with 6 bits μlog PCM (66.15kbits/s). We are convinced that coding method of LSP residual signals using wavelet analysis is an effective approach to synthesize speech.

Synthesis of Arabic from short sound clusters

This article describes an unrestricted vocabulary text-to-speech (TTS) conversion system for the synthesis of Standard Arabic (SA) speech. The system uses short phonetic clusters that are derived from the Arabic syllables to synthesize Arabic. Basic and phonetic variants of the synthesis units are defined after qualitative and quantitative analyses of the phonetics of SA. A speech database of the synthesis units and their phonetic variations is created and the units are tested to control their segmental quality. Besides the types of synthesis unit used, their enhancement with phonetic variants, and their segmental quality control, the production of good quality speech also depends on waveform analysis and the method used to concatenate the synthesis units together. Waveform analysis is needed to condition the selected synthesis units at their junctures to produce synthesized speech of better quality. The types of speech juncture between contiguous units, the phonetic characteristics of the sounds surrounding the junctures and the concatenation artifacts occurring across the junctures are important and will be discussed. The results of waveform analysis and smoothing algorithms will be presented. The intelligibility of synthesized Arabic by a standard intelligibility test method that is adapted to suit the Arabic phonetic characteristics and scoring the results of the tests will also be dealt with.

Speech enhancement for an in-ear communication system using bone-conducted and acoustic signals

The main requirement for mobile radiotelephony nowadays is an ergonomic, robust, and functional system. One weak point of existing devices, such as hands-free kits for mobile telephones, is poor speech intelligibility due to factors such as environmental noise and poor signal reception. This work investigates the potential of speech enhancement using post-processing of signals from two types of transducers: a miniature microphone placed on the opening of the ear canal, and a miniature accelerometer mounted on four locations around the head. The microphone provides good speech quality with the drawback of being sensitive to external noise, while the accelerometer senses vibrations that are immune to ambient noise, but are a distorted, low-pass-filtered version of speech sounds. The use of optimal filtering, which enhances the signals acquired from the miniature transducers by matching them to a reference recording of the same material, was investigated. The methods investigated here could then be applied to improve speech quality and reduce noise in communication earsets. Initial results with speech, based on sustained phonemes and phonetically balanced sentences, showed potential for improvement in the quality of the speech signals and also suggested that different filters might be used for different phonemic contexts. [Research funded by EPSRC Grant No. GR/M86026.]

A packet-based CAPDM speech coder for PCN applications

In this paper, we present a median-rate speech coder, the controlled adaptive prediction delta modulation coder (CAPDM), which operates at 16 kb/s with good speech quality and low algorithm complexity. The coder is dedicated to personal communication network (PCN) applications and transmits speech samples on the basis of packets. It combines the features of a one-step looking forward decision, syllabic companding, instantaneous companding, and adaptive prediction. In addition to the use of a short-term prediction filter, CAPDM also exploits the pitch property to predict speech waveform explicitly. With the aid of a pitch prediction filter, the performance of a CAPDM codec improves about 3 dB in segmental signal-to-noise ratio (SEGSNR). The average SEGSNR of CAPDM.FF is about 21 dB, which is 7 dB over traditional CVSD at 16 kb/s. We also utilize an adaptive postfilter (APF) to enhance the perceptual quality of the decoded speech. The mean opinion score (MOS) listening test of CAPDM.FF with APF shows that its average score achieves 4.19, which is as good as G.728 16-kb/s LD-CELP and is comparable with CCITT G.721 32-kb/s ADPCM. The complexity of CAPDM.FF is evaluated to be 8 MIPS, which is much lower than that of LD-CELP and could be further reduced by adopting a smaller correlation window for pitch detection. To solve the problem of packet loss, we developed a packet-based waveform substitution method by reinitializing the codec parameters at the beginning of each packet. The simulation results show that CAPDM.FF could tolerate 5% of packet loss and still keep an SEGSNR at 10 dB and an MOS at about 3.0.

Acoustic echo canceler

The invention provides an acoustic echo canceler which realizes sufficient echo suppression processing on an auditory sense which matches with an amount of calculation and realizes a good speech quality. The acoustic echo canceler includes a first analysis filter bank for dividing a reception signal into a plurality of sub-bands, a plurality of adaptive filters for individually producing false echo signals from the reception signal after divided into the sub-bands, a second analysis filter bank for dividing an input signal into a plurality of sub-bands, a plurality of subtractors for subtracting the false echo signals corresponding to the sub-bands from the input signal after divided into the sub-bands, and a synthesis filter bank for composing outputs of the subtractors. The acoustic echo canceler further includes a band information calculation section for estimating influences of the echoes for the individual sub-bands and determining a priority order based on the estimation. Each of the adaptive filters determines processing to be executed thereby based on the priority order.

System for wideband speech coding based on recursive filterbanks

Wideband speech coding becomes increasingly important because of the enhanced quality due to the extended frequency range. Possible applications are in various fields partly with fixed and partly with variable data rates: synchronous and asynchronous transmission over digital networks, mobile telephony, and multimedia systems. In this context a system is presented which is well suited for dialog communication. The signal analysis and synthesis is done by filterbanks based on recursive filters. The use of recursive filters implies short system delays in the filterbank. However, phase distortions are present and have to be considered carefully. Different methods were implemented to reduce phase distortions and to compensate group delay peaks. Coding is based on redundance and irrelevance reduction. The irrelevance reduction is implemented by the exploitation of psychoacoustical masking effects. Encoder and decoder are running on a single DSP32C signal processor in real time. The achieved data rate for good speech quality is 32 kbit/s and below.