A limited feedback time-delay neural network

Continuous density hidden Markov models (CD-HMMs) are an essential component of modern systems for automatic speech recognition (ASR). These models assign probabilities to the sequences of acoustic feature vectors extracted by signal processing of speech waveforms. In this chapter, we investigate a new framework for parameter estimation in CD-HMMs. Our framework is inspired by recent parallel trends in the fields of ASR and machine learning. In ASR, significant improvements in performance have been obtained by discriminative training of acoustic models. In machine learning, significant improvements in performance have been obtained by discriminative training of large margin classifiers. Building on both these lines of work, we show how to train CD-HMMs by maximizing an appropriately defined margin between correct and incorrect decodings of speech waveforms. We start by defining an objective function over a transformed parameter space for CD-HMMs, then describe how it can be optimized efficiently by simple gradient-based methods. Within this framework, we obtain highly competitive results for phonetic recognition on the TIMIT speech corpus. We also compare our framework for large margin training to other popular frameworks for discriminative training of CD-HMMs.

In this paper, an evaluation of various discriminant neural networks classifiers for text- independent speaker verification problem is presented. Each person to be verified has a personalized neural network model. A new classifier called neural tree network (NTN) is also examined for this application. The memoryless feedforward neural network architecture makes decisions based on static features. Time delay neural network (TDNNs) have proved to be an efficient way to handle the dynamic nature of speech. Furthermore, a model called recurrent time delay neural networks (RTDNNs), obtained through a local feedback connection at the first hidden layer level of TDNNs is investigated. The training is carried out by backpropagation for sequence algorithm. The database used is a subset of the TIMIT database consisting of 38 speakers from the same dialect region. The NTN is compared with the MLP, TDNN, and RTDNN. It is shown that NTN is found to perform better than the other neural networks classifiers. Also, a little bit performance improvement was achieved due to the addition of temporal information for text-independent speaker verification problem using TDNNs and RTDNNs. Finally, we described the experimental results obtained using different neural network models.© (1994) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

This paper investigated the speaker independent vowel recognition for Malay children using the Time Delay Neural Network (TDNN). Due to less research done on the children speech recognition, the temporal structure of the children speech was not fully understood. This 2 hidden layers TDNN was proposed to discriminate 6 Malay vowels: /a/, /e/, /ə/, /i/, /o/ and /u/. The speech database consisted of vowel sounds from 360 children speakers. Cepstral coefficient was normalized for the input of TDNN. The frame rate of the TDNN was tested with 10ms, 20ms, and 30ms. It was found out that the 30ms frame rate produced the highest vowel recognition accuracy with 81.92%. The TDNN also showed higher speech recognition rate compared to the previous studies that used Multilayer Perceptron.

Recently, x-vector [1] has been a successful and popular approach for speaker verification, which employs a time delay neural network (TDNN) and statistics pooling to extract speaker characterizing embedding from variable-length utterances. Improvement upon the x-vector has been an active research area, and enormous neural networks have been elaborately designed based on the x-vector, e.g., extended TDNN (E-TDNN) [2], factorized TDNN (F-TDNN) [3], and densely connected TDNN (D-TDNN) [4]. In this work, we try to identify the optimal architectures from a TDNN based search space employing neural architecture search (NAS), named SpeechNAS. Leveraging the recent advances in the speaker recognition, such as high-order statistics pooling, multi-branch mechanism, D-TDNN and angular additive margin softmax (AAM) loss with a minimum hyper-spherical energy (MHE), SpeechNAS automatically discovers five network architectures, from SpeechNAS-1 to SpeechNAS-5, of various numbers of parameters and GFLOPs on the large-scale text-independent speaker recognition dataset VoxCelebl. Our derived best neural network achieves an equal error rate (EER) of 1.02% on the standard test set of VoxCelebl, which surpasses previous TDNN based state-of-the-art approaches by a large margin.

This paper proposes a large‐vocabulary speech recognition system using a phoneme spotting method by a time‐delay neural network (TDNN) and a predictive LR parser. This is the first attempt to recognize large vocabulary speech using neural networks. The prediction of phonemes in words is performed by a predictive LR parser. Time alignment between predicted phonemes by the predicted LR parser and phoneme spotting results via TDNN is realized using a DTW (dynamic time warping) method. Speaker‐dependent recognition for a 5240‐word vocabulary using 2620 test words uttered by a male announcer resulted in a rate of 92.6 percent for the top choices, rates of 97.6 and 99.1 percent for the second and fifth choices, respectively.

A method for ankle torque prediction ahead of the current time is proposed in this paper. The mean average value of EMG signals from four muscles, alongside the joint angle and angular velocity of the right ankle, were used as input parameters to train a time-delayed artificial neural network. Data collected from five healthy subjects were used to generate the dataset to train and test the model. The model predicted ankle torque for five different future times from zero to 2 seconds. Model predictions were compared to torque calculated from inverse dynamics for each subject. The model predicted ankle torque up to 1 second ahead of time with normalized root mean squared error of less than 15 percent while the coefficient of determination was over 0.85.Clinical Relevance- the potential of the model for predicting joint torque ahead of time is helpful to establish an intuitive interaction between human and assistive robots. This model has application to assist patients with neurological disorders.

Sea state forecasting is of primary importance for marine operations related to offshore oil&gas industry. Operative decisions are based on marine weather forecasts. For this reason, at least once a day a bulletin of marine weather forecasts is delivered on board. Bulletins give a very good general overview of weather for the next days, while sometime they are not accurate relatively to the forecasted seastate values or their time of occurrence. In this work has been developed an Artificial Intelligence model named time delayed neural networks (TDNN) to improve the accuracy of short term sea state forecasting, combining bulletins forecasts with buoy measurements. The TDNN parameters are optimized by Genetic Algorithms. It has been found that the optimized TDNN produces very good forecasts, always outperforming bulletins, at least up to 24 hours ahead. Two configurations of TDNN are analyzed: in the first one, TDNN is trained with only buoy measurements, while in the second one TDNN is trained with both buoy and bulletin data. For 24 hours lead time, the combined use of buoy measurements and bulletins as input has been demonstrated to give better results than the use of only buoy measurements.

Prediction of hourly solar radiation using a novel hybrid model of ARMA and TDNN

The work is devoted to resolving the contradiction between the accuracy of modelling nonlinear dynamics and the speed of model construction under conditions of limited computing resources. The purposeof the work is to reduce the time for building time delay neural networks while ensuring a given accuracy in the tasks of identifying nonlinear dynamic objects with continuous characteristics. This goal is achieved by developing a method for pre-training neural networks that reflect the basic characteristics of the subject area. The scientific novelty of the work is the development of a method for identifying nonlinear dynamic objectsin the form of time delay neural networks based on the use of a set of basic pre-trained neural networks that reflect the typical properties of the subject area. In contrast to the traditional approach to pre-training, the developed method allows building models of lower complexity.A formal criterion is proposed for determining the moment of termination of the neural network pre-training, the use of which allows avoiding retraining of the base model and ensuring a significant reduction in the model training time on the target data set. The practical utility of the work lies in the development of an algorithm for the method of pre-training time delay neural networks in the tasks of identifying nonlinear dynamic objects with continuous characteristics, which allows to significantlyreduce the training time of neural networks without losing model accuracy. The value of this study is todetermine the area of effective use of the proposed method, namely, when the general and target datasets do not have significant differences and the target dataset is of sufficient size to reflect the properties of the research object.

In this paper, we present a model for Turkish speech recognition. The model is syllable-based, where the recognition is performed through syllables as speech recognition units. The main goal of the model is to recognize as much as possible of a given continuous speech by identifying only a small set of syllables in the language. For that purpose, only the syllable types with a higher frequency are selected for the recognition. The use of longer recognition units in speech recognition systems increases the success of the recognition since it is easier to detect the endpoints of syllables when compared to phonemes. On the other side, word-based recognition requires a very large dataset that includes all the words and word forms in the language, which is also another challenge. Hereby, we take the advantage of Turkish being an ortographically transparent and syllabified language. Our model employs time delay neural networks (TDNNs) for learning syllables. We achieve an accuracy of %65.6 on our large vocabulary continuous speech corpus. In addition, we define an algorithm for the automatic detection of syllable boundaries which gives an accuracy of %44. The automatic syllable boundary detection module is used for the recognition of isolated syllables rather than a continuous speech.

Rubber is the most significant and widely used material in the industry. Sri Lanka is well renowned for production of quality rubber. This study is undertaken to forecast natural rubber (NR) prices for upcoming years in Sri Lanka. Since rubber is a storable intermediate good, current population heavily depends on future prices. Before the year 2011, rubber has good price scale, but due to the lack of government intervention the rubber price has been decreased after 2011. With change of rubber price, the accurate forecast is extremely important in executing policies and making decisions for the future of rubber industry. Moreover, there is no research studies can be found which attempts to forecast monthly NR prices in Sri Lanka by using machine learning techniques. Monthly rubber prices from June 2005 to January 2019 was considered for this study. 80% of the data was used to fit the model and the rest was used for model performance evaluation. All the unit root tests confirmed that the first difference of log series was stationary. ARIMA (1,1,1) model was selected as the best model with lower Akaike Information Criterion (AIC) among the other candidate models which exhibits RMSE of 22.5039 and MAE of 16.6923. To find a better model which cater the instability of the NR prices properly, a dynamic machine learning technique, Time Delay Neural Network (TDNN) was employed. The architecture of the identified TDNN model with the hyper-parameter tunning consists of one hidden layer with sixteen neurons, and 1:16 time delays and exhibits lower errors: RMSE of 0.01347 and MAE of 0.0074. It can be concluded that TDNN perform better than the ARIMA model in forecasting NR prices Sri Lanka.

Cardiovascular pathologies cause 23.5% of human deaths, worldwide. An auto-diagnostic system monitoring heart activity, which can identify the early symptoms of cardiac illnesses, might reduce the death rate caused by these problems. Phonocardiography (PCG) is one of the possible techniques able to detect heart problems. Nevertheless, acoustic signal enhancement is required since it is exposed to various disturbances coming from different sources. The most common denoising enhancement is based on the Wavelet Transform (WT). However, the WT is highly susceptible to variations in the noise frequency distribution. This paper proposes a new adaptive denoising algorithm, which combines WT and Time Delay Neural Networks (TDNN). The acquired signal is decomposed by means of the WT using the coif five-wavelet basis at the tenth decomposition level and then provided as input to the TDNN. Besides the advantage of adaptive thresholding, the reason for using TDNNs is their capacity of estimating the Inverse Wavelet Transform (IWT). The best parameters of the TDNN were found for a NN consisting of 25 neurons in the first and 15 in the second layer and the delay block of 12 samples. The method was evaluated on several pathological heart sounds and on signals recorded in a noisy environment. The performance of the developed system with respect to other wavelet-based denoising approaches was validated by the online questionnaire.

Time-delayed neural networks (TDNNs) can be used to learn the dynamics of an unknown system from input-output data. In many cases a model of the system is also available in the form of a system of ODES, derived either from first principles or using heuristic arguments. In such cases a functional comparison can be made between the dynamic behaviour of the model with that of the trained TDNN for arbitrary inputs. We show that Volterra kernels for both the model system and the TDNN can be obtained, and thus the system responses compared, in a manner that is independent of the input. The techniques of structural bilinearisation of ODES and Volterra series expansion of a TDNN, are demonstrated by application to the Hodgkin-Huxley set of equations.

Presented in this work is a possible myoelectric control scheme for a rehabilitation robotic application. The control input is from a time delayed neural network (TDNN). The input to the TDNN is four electromyographic (EMG) signals associated with the movement of the elbow and shoulder joints. The output of the TDNN is the joint position of the elbow and the joint position of the shoulder in the sagittal plane. The results presented here show the possibility of controlling multiple degrees of freedom at once. Prior work has shown that the optimal delay for accurate position prediction from a TDNN was 875ms with a 125ms interval, but this work shows that a delay of 300ms and a 100ms interval achieves similar results. This points to the feasibility of a TDNN based control scheme.

Time-delay recurrent neural network for temporal correlations and prediction