| Total: 116
In order to allocate the healthcare resource, triage classification system plays an important role in assessing the severity of illness of the boarding patient at emergency department. The self-report pain intensity numerical-rating scale (NRS) is one of the major modifiers of the current triage system based on the Taiwan Triage and Acuity Scale (TTAS). The validity and reliability of self-report scheme for pain level assessment is a major concern. In this study, we model the observed expressive behaviors, i.e., facial expressions and vocal characteristics, directly from audio-video recordings in order to measure pain level for patients during triage. This work demonstrates a feasible model, which achieves an accuracy of 72.3% and 51.6% in a binary and ternary pain intensity classification. Moreover, the study result reveals a significant association of current model and analgesic prescription/patient disposition after adjusted for patient-report NRS and triage vital signs.
Acoustical analysis of speech is considered a favorable and promising approach to objective assessment of voice disorders. Previous research emphasized on the extraction and classification of voice quality features from sustained vowel sounds. In this paper, an investigation on voice assessment using continuous speech utterances of Cantonese is presented. A DNN-HMM based speech recognition system is trained with speech data of unimpaired voice. The recognition accuracy for pathological utterances is found to decrease significantly with the disorder severity increasing. Average acoustic posterior probabilities are computed for individual phones from the speech recognition output lattices and the DNN soft-max layer. The phone posteriors obtained for continuous speech from the mild, moderate and severe categories are highly distinctive and thus useful to the determination of voice disorder severity. A subset of Cantonese phonemes are identified to be suitable and reliable for voice assessment with continuous speech.
Long-term voice outcomes of 13 tracheoesophageal speakers are assessed using speech samples that were recorded with at least 7 years in between. Intelligibility and voice quality are perceptually evaluated by 10 experienced speech and language pathologists. In addition, automatic speech evaluations are performed with tools from Ghent University. No significant group effect was found for changes in voice quality and intelligibility. The recordings showed a wide interspeaker variability. It is concluded that intelligibility and voice quality of tracheoesophageal voice is mostly stable over a period of 7 to 18 years.
Mild Cognitive Impairment (MCI), sometimes regarded as a prodromal stage of Alzheimer’s disease, is a mental disorder that is difficult to diagnose. Recent studies reported that MCI causes slight changes in the speech of the patient. Our previous studies showed that MCI can be efficiently classified by machine learning methods such as Support-Vector Machines and Random Forest, using features describing the amount of pause in the spontaneous speech of the subject. Furthermore, as hesitation is the most important indicator of MCI, we took special care when handling filled pauses, which usually correspond to hesitation. In contrast to our previous studies which employed manually constructed feature sets, we now employ (automatic) correlation-based feature selection methods to find the relevant feature subset for MCI classification. By analyzing the selected feature subsets we also show that features related to filled pauses are useful for MCI detection from speech samples.
Approximately 60% of children with speech and language impairments do not receive the intervention they need because their impairment was missed by parents and professionals who lack specialized training. Diagnoses of these disorders require a time-intensive battery of assessments, and these are often only administered after parents, doctors, or teachers show concern. An automated test could enable more widespread screening for speech and language impairments. To build classification models to distinguish children with speech or language impairments from typically developing children, we use acoustic features describing speech and pause events in story retell tasks. We developed and evaluated our method using two datasets. The smaller dataset contains many children with severe speech or language impairments and few typically developing children. The larger dataset contains primarily typically developing children. In three out of five classification tasks, even after accounting for age, gender, and dataset differences, our models achieve good discrimination performance (AUC > 0.70).
The quality of cleft lip and palate (CLP) speech is affected due to hyper-nasality and mis-articulation. Surgery and speech therapy are required to correct the structural and functional defects of CLP, which will result in an enhanced speech signal. The quality of the enhanced speech is perceptually evaluated by speech-language pathologists and results are highly biased. In this work, a signal processing based two stage speech enhancement method is proposed to get the perceptual benchmark to compare the signal after the surgery / therapy. In the first stage, CLP speech is enhanced by suppressing the nasal formant and in the second stage, spectral peak-valley enhancement is carried out to reduce the hyper-nasality associated with the CLP speech. The evaluation results show that the perceptual quality of CLP speech signal is improved after enhancement in both stages. Further, the improvement in the quality of the enhanced signal is compared with the speech signal after palatal prosthesis / surgery. The perceptual evaluation results show that the enhanced speech signals are better than the speech after prosthesis / surgery
Most existing single-channel noise-suppression algorithms cannot improve speech intelligibility for normal-hearing listeners; however, the underlying reason for this performance deficit is still unclear. Given that various speech segments contain different perceptual contributions, the present work assesses whether the intelligibility of noisy speech can be improved when selectively suppressing its noise at high-level (vowel-dominated) or middle-level (containing vowel-consonant transitions) segments by existing single-channel noise-suppression algorithms. The speech signal was corrupted by speech-spectrum shaped noise and two-talker babble masker, and its noisy high- or middle-level segments were replaced by their noise-suppressed versions processed by four types of existing single-channel noise-suppression algorithms. Experimental results showed that performing segmental noise-suppression at high- or middle-level led to decreased intelligibility relative to noisy speech. This suggests that the lack of intelligibility improvement by existing noise-suppression algorithms is also present at segmental level, which may account for the deficit traditionally observed at full-sentence level.
Most recently proposed near-end speech enhancement methods have been evaluated with the overall power (RMS) of the speech held constant. While significant intelligibility gains have been reported in various noisy conditions, an equal-RMS constraint may lead to enhancement solutions that increase the loudness of the original speech. Comparable effects might be produced simply by increasing the power of the original speech, which also leads to an increase in loudness. Here we suggest modifying the equal-RMS constraint to one of equal loudness between the original and the modified signals, based on a loudness model for time-varying sounds. Four state-of-the-art speech-in-noise intelligibility enhancement systems were evaluated under the equal-loudness constraint, using intelligibility tests with normal-hearing listeners. Results were compared with those obtained under the equal-RMS constraint. The methods based on spectral shaping and dynamic range compression yielded significant intelligibility gains regardless of the constraint, while for the method without dynamic range compression the intelligibility gain was lower under the equal-loudness than under the equal-RMS constraint.
Many studies have shown the advantage of using ideal binary masking (IdBM) to improve the intelligibility of speech corrupted by interfering maskers. Given the fact that amplitude and phase are two important acoustic cues for speech perception, the present work further investigated the relative contributions of these two cues to the intelligibility advantage of IdBM-processed sentences. Three types of Mandarin IdBM-processed stimuli (i.e., amplitude-only, phase-only, and amplitude-and-phase) were generated, and played to normal-hearing listeners to recognize. Experiment results showed that amplitude- or phase-only cue could lead to significantly improved intelligibility of IdBM-processed sentences in relative to noise-masked sentences. A masker-dependent amplitude over phase advantage was observed when accounting for their relative contributions to the intelligibility advantage of IdBM-processed sentences. Under steady-state speech-spectrum shaped noise, both amplitude- and phase-only IdBM-processed sentences contained intelligibility information close to that contained in amplitude-and-phase IdBM-processed sentences. In contrast, under competing babble masker, amplitude-only IdBM-processed sentences were more intelligible than phase-only IdBM-processed sentences, and neither could account for the intelligibility advantage of amplitude-and-phase IdBM-processed sentences.
State-of-the-art binaural objective intelligibility measures (OIMs) require individual source signals for making intelligibility predictions, limiting their usability in real-time online operations. This limitation may be addressed by a blind source separation (BSS) process, which is able to extract the underlying sources from a mixture. In this study, a speech source is presented with either a stationary noise masker or a fluctuating noise masker whose azimuth varies in a horizontal plane, at two speech-to-noise ratios (SNRs). Three binaural OIMs are used to predict speech intelligibility from the signals separated by a BSS algorithm. The model predictions are compared with listeners’ word identification rate in a perceptual listening experiment. The results suggest that with SNR compensation to the BSS-separated speech signal, the OIMs can maintain their predictive power for individual maskers compared to their performance measured from the direct signals. It also reveals that the errors in SNR between the estimated signals are not the only factors that decrease the predictive accuracy of the OIMs with the separated signals. Artefacts or distortions on the estimated signals caused by the BSS algorithm may also be concerns.
Speech intelligibility in reverberant environments is reduced because of overlap-masking. Signal modification prior to presentation in such listening environments, e.g., with a public announcement system, can be employed to alleviate this problem. Time-scale modifications are particularly effective in reducing the effect of overlap-masking. A method for introducing linguistically-motivated pauses is proposed in this paper. Given the transcription of a sentence, pause strengths are predicted at word boundaries. Pause duration is obtained by combining the pause strength and the time it takes late reverberation to decay to a level where a target signal-to-late-reverberation ratio criterion is satisfied. Considering a moderate reverberation condition and both binary and continuous pause strengths, a formal listening test was performed. The results show that the proposed methodology offers a significant intelligibility improvement over unmodified speech while continuous pause strengths offer an advantage over binary pause strengths.
Speech and audio codecs model the overall shape of the signal spectrum using envelope models. In speech coding the predominant approach is linear predictive coding, which offers high coding efficiency at the cost of computational complexity and a rigid systems design. Audio codecs are usually based on scale factor bands, whose calculation and coding is simple, but whose coding efficiency is lower than that of linear prediction. In the current work we propose an entropy coding approach for scale factor bands, with the objective of reaching the same coding efficiency as linear prediction, but simultaneously retaining a low computational complexity. The proposed method is based on quantizing the distribution of spectral mass using beta-distributions. Our experiments show that the perceptual quality achieved with the proposed method is similar to that of linear predictive models with the same bit rate, while the design simultaneously allows variable bit-rate coding and can easily be scaled to different sampling rates. The algorithmic complexity of the proposed method is less than one third of traditional multi-stage vector quantization of linear predictive envelopes.
In this paper we introduce an objective evaluation methodology for Blind Bandwidth Extension (BBE) algorithms. The methodology combines an objective method, POLQA, with a bandwidth requirement, based on a frequency mask. We compare its results to subjective test data, and show that it gives consistent results across several bandwidth extension algorithms. Additionally, we show that our latest BBE algorithm achieves quality similar to AMR-WB at 8.85 kbps, using both subjective and objective evaluation methods.
This paper discusses the voice and audio quality characteristics of the EVS, the recently standardized 3GPP Enhanced Voice Services codec. Especially frame erasure conditions with and without channel aware mode were evaluated. The test consisted of two extended range MOS listening tests. The tests contained both clean and noisy speech in clean channel as well as with four frame erasure rates (5%, 10%, 20% and 30%) for selected codecs and bitrates. In addition to subjective test results some additional objective results are presented. The results show that EVS channel aware mode performs better than EVS native mode in high FER rates. For comparison also AMR, AMR-WB and Opus codecs were included to the listening tests.
In this work, we describe an interaural magnification algorithm for speech enhancement in noise and reverberation. The proposed algorithm operates by magnifying the interaural level differences corresponding to the interfering sound source. The enhanced signal outputs are estimated by processing the signal inputs with the interaurally-magnified head-related transfer functions. Experimental results with speech masked by a single interfering source in anechoic and reverberant scenarios indicate that the proposed algorithm yields an increased benefit due to spatial release from masking and a much higher perceived speech quality.
Speech recognition in multi-channel environments requires target speaker localization, multi-channel signal enhancement and robust speech recognition. We here propose a system that addresses these problems: Localization is performed with a recently introduced probabilistic localization method that is based on support-vector machine learning of GCC-PHAT weights and that estimates a spatial source probability map. The main contribution of the present work is the introduction of a probabilistic approach to (re-)estimation of location-specific steering vectors based on weighting of observed inter-channel phase differences with the spatial source probability map derived in the localization step. Subsequent speech recognition is carried out with a DNN-HMM system using amplitude modulation filter bank (AMFB) acoustic features which are robust to spectral distortions introduced during spatial filtering. The system has been evaluated on the CHIME-3 multi-channel ASR dataset. Recognition was carried out with and without probabilistic steering vector re-estimation and with MVDR and delay-and-sum beamforming, respectively. Results indicate that the system attains on real-world evaluation data a relative improvement of 31.98% over the baseline and of 21.44% over a modified baseline. We note that this improvement is achieved without exploiting oracle knowledge about speech/non-speech intervals for noise covariance estimation (which is, however, assumed for baseline processing).
In this paper, a priori speech absence probability (SAP) estimator is proposed for accurately obtaining the speech presence probability (SPP) in a complex noise field. Unlike previous techniques, the proposed estimator considers a complex noise sound field where the target speech is corrupted by a coherent interference with diffuse noise around. The proposed algorithm estimates a priori SAP based on the normalized speech to interference plus diffuse noise ratio (SINR) being expressed in terms of the speech to interference ratio (SIR) and the directional to diffuse noise ratio (DDR). The SIR is obtained from a quadratic equation of the magnitude-squared coherence (MSC) between two microphone signals. A performance comparison with several advanced a priori SAP estimators was conducted in terms of the receiver operating characteristic (ROC) curve. The proposed algorithm attains a correct detection rate at a given false-alarm rate that is higher than those attained by conventional algorithms.
Psychoacoustic active noise control (ANC) systems by considering human hearing thresholds in different frequency bands were developed in the past. Besides the frequency sensitivity, human hearing also shows different sensitivity to spectral and temporal modulations of the sound. In this paper, we propose a new psychoacoustic active noise control system by further considering the spectral modulation sensitivity of human hearing. In addition to the sound pressure level (SPL), the loudness level is also objectively assessed to evaluate the noise reduction performance of the proposed ANC system. Simulation results demonstrate the proposed system outperforms two compared systems under test conditions of narrowband and broadband noise in terms of the loudness level. The proposed algorithm has been validated on TI C6713 DSP platform for real-time process.
The work presented in this paper is motivated by our earlier work exploring sparse representation based approach for automatic bandwidth extension (ABWE) of speech signals. In that work, two dictionaries one for voiced and the other for unvoiced speech frames are created using KSVD algorithm on wideband data. Each of the atoms of these dictionaries is then decimated and interpolated by a factor of 2 to generate narrowband interpolated (NBI) dictionaries whose atoms have one-to-one correspondence with those of the WB dictionaries. The given narrowband speech frames are also interpolated to generated NBI targets and those are sparse coded over the NBI dictionaries. The resulting sparse codes are then applied to the WB dictionaries to estimate the WB target data. In this work, we extend the said approach by making use of an existing semi-coupled dictionary learning (SCDL) algorithm. Unlike the direct dictionary learning, the SCDL algorithm also learns a set of bidirectional transforms coupling the dictionaries more flexibly. The bandwidth enhanced speech obtained employing the SCDL approach and a modified high/low band gain adjustment yields significant improvements in terms of speech quality measures as well as in the context of children’s mismatched speech recognition.
Neurological disorders disrupt brain functions, affecting the life of many individuals. Conventional neurological disorder diagnosis methods require inconvenient and expensive devices. Several studies have identified speech biomarkers that are informative of neurological disorders, so speech-based interfaces can provide effective, convenient and affordable prescreening tools for diagnosis. We have investigated stand-alone automatic speech-based assessment tools for portable devices. Our current data collection protocol includes seven brief tests for which we have developed specialized automatic speech recognition (ASR) systems. The most challenging task from an ASR perspective is a popular diadochokinetic test consisting of fast repetitions of “PA-TA-KA”, where subjects tend to alter, replace, insert or skip syllables. This paper presents our efforts to build a speech-based application specific for this task, where the computation is fast, efficient, and accurate on a portable device, not in the cloud. The tool recognizes the target syllables, providing phonetic alignment. This information is crucial to reliably estimate biomarkers such as the number of repetitions, insertions, mispronunciations, and temporal prosodic structure of the repetitions. We train and evaluate the application for two neurological disorders: traumatic brain injuries (TBIs) and Parkinson’s disease. The results show low syllable error rates and high boundary detection, across populations.
This paper is focused on the application of the Language Identification (LID) technology for intelligent vehicles. We cope with short sentences or words spoken in moving cars in four languages: English, Spanish, German, and Finnish. As the response time of the LID system is crucial for user acceptance in this particular task, speech signals of different durations with total average of 3.8s are analyzed. In this paper, the authors propose the use of Deep Neural Networks (DNN) to model effectively the i-vector space of languages. Both raw i-vectors and session variability compensated i-vectors are evaluated as input vectors to DNNs. The performance of the proposed DNN architecture is compared with both conventional GMM-UBM and i-vector/LDA systems considering the effect of durations of signals. It is shown that the signals with durations between 2 and 3s meet the requirements of this application, i.e., high accuracy and fast decision, in which the proposed DNN architecture outperforms GMM-UBM and i-vector/LDA systems by 37% and 28%, respectively.
i-vector modeling techniques have been successfully used for speaker clustering task recently. In this work, we propose the extraction of i-vectors from short- and long-term speech features, and the fusion of their PLDA scores within the frame of speaker diarization. Two sets of i-vectors are first extracted from short-term spectral and long-term voice-quality, prosodic and glottal to noise excitation ratio (GNE) features. Then, the PLDA scores of these two i-vectors are fused for speaker clustering task. Experiments have been carried out on single and multiple site scenario test sets of Augmented Multi-party Interaction (AMI) corpus. Experimental results show that i-vector based PLDA speaker clustering technique provides a significant diarization error rate (DER) improvement than GMM based BIC clustering technique.
Several functional models for head-related transfer function (HRTF) have been proposed based on spherical harmonic (SH) orthogonal functions, which yield an encouraging performance level in terms of log-spectral distortion (LSD). However, since the properties of subbands are quite different and highly subject-dependent, the degree of SH expansion should be adapted to the subband and the subject, which is quite challenging. In this paper, a sparse spherical harmonic-based model termed SSHM is proposed in order to achieve an intelligent frequency truncation. Different from SH-based model (SHM) which assigns the degree for each subband, SSHM constrains the number of SH coefficients by using an l1 penalty, and automatically preserves the significant coefficients in each subband. As a result, SSHM requires less coefficients at the same SD level than other truncation methods to reconstruct HRTFs . Furthermore, when used for interpolation, SSHM gives a better fitting precision since it naturally reduces the influence of the fluctuation caused by the movement of the subject and the processing error. The experiments show that even using about 40% less coefficients, SSHM has a slightly lower LSD than SHM. Therefore, SSHM can achieve a better tradeoff between efficiency and accuracy.
Deep clustering is a recently introduced deep learning architecture that uses discriminatively trained embeddings as the basis for clustering. It was recently applied to spectrogram segmentation, resulting in impressive results on speaker-independent multi-speaker separation. In this paper we extend the baseline system with an end-to-end signal approximation objective that greatly improves performance on a challenging speech separation. We first significantly improve upon the baseline system performance by incorporating better regularization, larger temporal context, and a deeper architecture, culminating in an overall improvement in signal to distortion ratio (SDR) of 10.3 dB compared to the baseline of 6.0 dB for two-speaker separation, as well as a 7.1 dB SDR improvement for three-speaker separation. We then extend the model to incorporate an enhancement layer to refine the signal estimates, and perform end-to-end training through both the clustering and enhancement stages to maximize signal fidelity. We evaluate the results using automatic speech recognition. The new signal approximation objective, combined with end-to-end training, produces unprecedented performance, reducing the word error rate (WER) from 89.1% down to 30.8%. This represents a major advancement towards solving the cocktail party problem.