Large vocabulary isolated words recognizer on a cellular array processor

DTW algorithm compares the parameters of an unknown spoken word with the parameters of one or more reference templates. The more reference templates are used for the same word, the higher is the recognition rate. But increasing the number of reference templates for the same word to recognize, leads to an increase in memory resources and computing time. The proposed algorithm is used in the learning phase and combines the advantages of DTW and Vector Quantization (VQ); instead of storing multiple reference templates, it stores only one reference model for each word and that reference is based on classes (like in the vector quantization method), each class is represented by a centroid (or codeword). In the recognition phase, the parameters of the unknown utterance are compared to the centroids of the reference model. This solution increases the speed of calculation in the recognition phase and reduces the quantity of used memory.

In this report three front ends, a fixed Q cochlear filter (FQF), an adaptive Q cochlear filter (AQF), and a Bark DFT (DFT), are compared for use as the front end of a DTW system. The FQF is a conventional cascade/parallel-type cochlear filter that stimulates the asymmetrical filtering characteristics of a basilar membrane system. The AQF is a nonlinear cochlear filter that simulates three level-dependent characteristics of a basilar membrane system [T. Hirahara etal., Proc. ICASSP, 496–499 (1989)]. The DFT front end generates 64-channel Bark scale coefficients based on a 512-point DFT magnitude spectrum. These three front ends have 64 channels covering the frequency range from 1.5 to 19.5 Bark. Recognition performance for clean speech, degraded speech by adding noise and/or reverberation, and under multispeaker conditions, are compared. Four signal-to-noise ratios, S/N=∞, 40, 20, and 10 dB, are set by adding different levels of pink noise to speech data. As for reverberant speech, the impulse responses obtained in the ATR reverberation room, RT=0.2 and 1.1 s, are convolved with speech data. Speech data used in the experiments are 216 phoneme-balanced Japanese words uttered by two male and two female speakers. A standard dynamic time warping (DTW) system is used as a back end. The experiments results are as follows: (1) For noisy speech, AQF performance is equal to that of FQF but more robust than that of DFT. (2) For reverberant speech, AQF is affected more than DFT but the performance is better than that of FQF. (3) For speaker variation, AQF gives better performance than do FQF or DFT. While the advantage of the AQF front end is small with an HMM back end [T. Hirahara etal., Proc. ICSLP, 381–384 (1990)], these results show that the AQF is an excellent front end for a DTW recognition system.

Dynamic Time Warping (DTW) is a well-known similarity measure for time series. The standard dynamic programming approach to compute the DTW distance of two length-n time series, however, requires O(n^2) time, which is often too slow for real-world applications. Therefore, many heuristics have been proposed to speed up the DTW computation. These are often based on lower bounding techniques, approximating the DTW distance, or considering special input data such as binary or piecewise constant time series. In this paper, we present a first exact algorithm to compute the DTW distance of two run-length encoded time series whose running time only depends on the encoding lengths of the inputs. The worst-case running time is cubic in the encoding length. In experiments we show that our algorithm is indeed fast for time series with short encoding lengths.

VLSI architectures for string matching and pattern matching

Speech recognition system using dynamic time warping are considered. It is proposed to generate a single pattern to be used as a time aligning pattern (TAP) for all the words in the system (reference and unknowns). An unknown word is time-aligned using only the TAP, and the distance between the time-aligned version and any of the reference templates (which have already been time-aligned with the tap in the training mode) is directly calculated without any further need for dynamic time warping (DTW) operations. A great reduction in the processing time results. Two methods for generating TAPs are proposed. In both cases, the processing time savings, compared to DTW, is in the range of 95% for comparable recognition accuracy. >

Unmanned Arial Vehicles (UAVs) require the development of some on-board safety equipments before inheriting the sky. An on-board collision avoidance system is being built by our team. Due to the strict size, weight, power, and costs constraints, visual intruder airplane detection is the only option. This paper introduces our visual airplane detector algorithm, which is designed to be operational in clear and in cloudy situations under regular daylight visual conditions. To be able to implement the algorithm on-board, we have carefully selected topographic operators, which can be efficiently solved on cellular processor arrays.

Over the last decade, time series motif discovery has emerged as a useful primitive for many downstream analytical tasks, including clustering, classification, rule discovery, segmentation, and summarization. In parallel, there has been an increased understanding that Dynamic Time Warping (DTW) is the best time series similarity measure in a host of settings. Surprisingly however, there has been virtually no work on using DTW to discover motifs. The most obvious explanation of this is the fact that both motif discovery and the use of DTW can be computationally challenging, and the current best mechanisms to address their lethargy are mutually incompatible. In this work, we present the first scalable exact method to discover time series motifs under DTW. Our method automatically performs the best trade-off between time-to-compute and tightness-of-lower-bounds for a novel hierarchy of lower bounds representation we introduce. We show that under realistic settings, our algorithm can admissibly prune up to 99.99% of the DTW computations.

A probability decision criterion for speech recognition

Spectral estimation often forms the basis for distinguishing and tracking signals of interest in the presence of noise and for extracting information from the received data. The application of Fourier techniques to the problem of estimating the properties of sinusoids in noise dates back as far as Shuster (1898), Fourier spectrum analysis is the basis for almost all spectral-estimation equipment, * including the common sweeping-filter spectrum analyzer, the parallel filter bank, the fast Fourier transform (FFT), the delay-line time compressor (Deltic), and the compressive spectrum analyzer (Microscan). A problem with Fourier spectrum analysis, however, is that it makes implicit assumptions concerning data outside the observation interval and, frequently, these physically unrealistic assumptions reduce the quality of the estimates. During the past decade, two radically different non-Fourier spectral-estimation techniques have emerged — maximum-entropy spectrum analysis (Burg, 1967) and spectral decomposition (Pisarenko, 1973). These techniques offer alternative and often more realistic data models which, in many cases, lead to better estimation performance. This paper reviews the Fourier methods and compares them to the new techniques in terms of signal models assumed by the three basic methods and their ability to distinguish multiple sinusoids in noise.

This paper proposes the design of a novel current-mode front-end for the extraction of localization spectral cues from two audio signals, together with test results. The front-end consists of two parallel filter banks, envelope extraction and comparison circuitry, together with an AGC loop. The extracted cues are intended to be further processed in order to determine the source azimuth and elevation. A current-mode log-domain implementation using subthreshold MOS operation is used for micropower operation while still achieving a good bandwidth and linearity. A current-mode solution is also preferred because of the ease of implementation of certain mathematical operations. The front-end splits the input signals into different frequency bands and computes monaural and interaural spectral cues from the resulting signal envelopes for each band. The front-end has been optimized to operate at a supply voltage of 1.8 V and most blocks have been designed using a differential architecture. To our knowledge, this is the first log-domain implementation of a front-end for 2-D localization cues extraction. The design has been carried out using a standard double-poly double-metal 0.8 μm CMOS process with VT e 0.8 V. The bandpass filters which form the main core of the chip exhibit a measured dynamic range of 62 dB corresponding to 1.9% THD, while the total power dissipation is 890 μW.

Meditation with mobile apps has been shown to improve mental and physical health. However, regular, long-term meditation app use is needed to maintain these health benefits, and many people have a difficult time maintaining engagement with meditation apps over time. Our goal was to determine the length of the timeframe over which usage data must be collected before future app abandonment can be predicted accurately in order to better target additional behavioral support to those who are most likely to stop using the app. Data were collected from a randomly drawn sample of 2600 new subscribers to a 1-year membership of the mobile app Calm, who started using the app between July and November of 2018. App usage data contained the duration and start time of all meditation sessions with the app over 365 days. We used these data to construct the following predictive model features: total daily sessions, total daily duration, and a measure of temporal similarity between consecutive days based on the dynamic time warping (DTW) distance measure. We then fit random forest models using increasingly longer periods of data after users subscribed to Calm to predict whether they performed any meditation sessions over 2-week intervals in the future. Model fit was assessed using the area under the receiver operator characteristic curve (AUC), and an exponential growth model was used to determine the minimal amount of data needed to reach an accurate prediction (95% of max AUC) of future engagement. After first subscribing to Calm, 83.1% of the sample used the Calm app on at least 1 more day. However, by day 350 after subscribing, 58.0% of users abandoned their use of the app. For the persistent users, the average number of daily sessions was 0.33 (SD = 0.02), the average daily duration of meditating was 3.93 minutes (SD = 0.25), and the average DTW distance to the previous day was 1.50 (SD = 0.17). The exponential growth models revealed that an average of 64 days of observations after subscribing to Calm are needed to reach an accurate prediction of future app engagement. Our results are consistent with existing estimates of the time required to develop a new habit. Additionally, this research demonstrates how to use app usage data to quickly and accurately predict the likelihood of users' future app abandonment. This research allows future researchers to better target just-in-time interventions towards users at risk of abandonment.

Auditory nerve representation as a front-end for speech recognition in a noisy environment

ObjectiveElectroencephalogram (EEG) based brain–computer interfaces (BCI) in motor imagery (MI) have developed rapidly in recent years. A reliable feature extraction method is essential because of a low signal-to-noise ratio (SNR) and time-dependent covariates of EEG signals. Because of efficient application in various fields, deep learning has been adopted in EEG signal processing and has obtained competitive results compared with the traditional methods. However, designing and training an end-to-end network to fully extract potential features from EEG signals remains a challenge in MI.ApproachIn this study, we propose a parallel multiscale filter bank convolutional neural network (MSFBCNN) for MI classification. We introduce a layered end-to-end network structure, in which a feature-extraction network is used to extract temporal and spatial features. To enhance the transfer learning ability, we propose a network initialization and fine-tuning strategy to train an individual model for inter-subject classification on small datasets. We compare our MSFBCNN with the state-of-the-art approaches on open datasets.ResultsThe proposed method has a higher accuracy than the baselines in intra-subject classification. In addition, the transfer learning experiments indicate that our network can build an individual model and obtain acceptable results in inter-subject classification. The results suggest that the proposed network has superior performance, robustness, and transfer learning ability.

A method for time-frequency multirate filtering of discrete-time signals is proposed. An input signal is nonuniformly decomposed into a number of subsignals, and the signals may be modified and combined on a specified time-frequency plane in real time processing. The modification process implies adaptive signal enhancement, filtering, movement and exchange over the time-frequency plane. The multirate filter is constructed based on parallel analysis and synthesis filter bank structures, and time-varying filtering may be achieved. By introducing a time-frequency plane and bifrequency maps, we formulate the multirate filters design completely. Design examples and experimental simulations are included.

Designing filter banks for source coding purposes classically relies on perfect reconstruction filters. However, several studies have shown recently that taking the quantization noise into account in the design could yield noticeable reduction of the reconstruction distortion. In particular, a joint optimization of the synthesis filters and quantizers with respect to the output mean squared error (MSE) was proposed, and the resulting scheme was called an minimum MSE (MMSE) filter bank. However, this approach was dedicated to parallel uniform filter banks, which require a very large computational complexity. This paper is concerned with the possibility of providing MMSE filters under the constraint that the filter bank is either of a modulated or a tree-structured kind and thus of low implementation cost. Various approaches for the optimization of the synthesis bank are presented, taking into account the structural constraint. Depending on the number of parameters to be optimized, we obtain various tradeoffs between decoder complexity, transmitted bit rate, and reconstruction distortion.