A G Ramakrishnan

Anoop C S, Prathosh A P, A G Ramakrishnan

Building an automatic speech recognition (ASR) system from scratch requires a large amount of annotated speech data, which is difficult to collect in many languages. However, there are cases where the low-resource language shares a common acoustic space with a high-resource language having enough annotated data to build an ASR. In such cases, we show that the domain-independent acoustic models learned from the high-resource language through unsupervised domain adaptation (UDA) schemes can enhance the performance of the ASR in the low-resource language. We use the specific example of Hindi in the source domain and Sanskrit in the target domain. We explore two architectures: i) domain adversarial training using gradient reversal layer (GRL) and ii) domain separation networks (DSN). The GRL and DSN architectures give absolute improvements of 6.71% and 7.32%, respectively, in word error rate over the baseline deep neural network model when trained on just 5.5 hours of data in the target domain. We also show that choosing a proper language (Telugu) in the source domain can bring further improvement. The results suggest that UDA schemes can be helpful in the development of ASR systems for low-resource languages, mitigating the hassle of collecting large amounts of annotated speech data.

Ram Krishna Pandey, Souvik Karmakar, A G Ramakrishnan et al.

In this work, we have proposed several enhancements to improve the performance of any facial emotion recognition (FER) system. We believe that the changes in the positions of the fiducial points and the intensities capture the crucial information regarding the emotion of a face image. We propose the use of the gradient and the Laplacian of the input image together with the original input into a convolutional neural network (CNN). These modifications help the network learn additional information from the gradient and Laplacian of the images. However, the plain CNN is not able to extract this information from the raw images. We have performed a number of experiments on two well known datasets KDEF and FERplus. Our approach enhances the already high performance of state-of-the-art FER systems by 3 to 5%.

Ram Krishna Pandey, K Vignesh, A G Ramakrishnan et al.

There is a need for information retrieval from large collections of low-resolution (LR) binary document images, which can be found in digital libraries across the world, where the high-resolution (HR) counterpart is not available. This gives rise to the problem of binary document image super-resolution (BDISR). The objective of this paper is to address the interesting and challenging problem of super resolution of binary Tamil document images for improved readability and better optical character recognition (OCR). We propose multiple deep neural network architectures to address this problem and analyze their performance. The proposed models are all single image super-resolution techniques, which learn a generalized spatial correspondence between the LR and HR binary document images. We employ convolutional layers for feature extraction followed by transposed convolution and sub-pixel convolution layers for upscaling the features. Since the outputs of the neural networks are gray scale, we utilize the advantage of power law transformation as a post-processing technique to improve the character level pixel connectivity. The performance of our models is evaluated by comparing the OCR accuracies and the mean opinion scores given by human evaluators on LR images and the corresponding model-generated HR images.

Amit Meghanani, A G Ramakrishnan

We propose a new feature, namely, pitchsynchronous discrete cosine transform (PS-DCT), for the task of speaker identification. These features are obtained directly from the voiced segments of the speech signal, without any preemphasis or windowing. The feature vectors are vector quantized, to create one separate codebook for each speaker during training. The performance of the PS-DCT features is shown to be good, and hence it can be used to supplement other features for the speaker identification task. Speaker identification is also performed using Mel-frequency cepstral coefficient (MFCC) features and combined with the proposed features to improve its performance. For this pilot study, 30 speakers (14 female and 16 male) have been picked up randomly from the TIMIT database for the speaker identification task. On this data, both the proposed features and MFCC give an identification accuracy of 90% and 96.7% for codebook sizes of 16 and 32, respectively, and the combined features achieve 100% performance. Apart from the speaker identification task, this work also shows the capability of DCT to capture discriminative information from the speech signal with minimal pre-processing.

Ram Krishna Pandey, Nabagata Saha, Samarjit Karmakar et al.

With the recent advancement in the deep learning technologies such as CNNs and GANs, there is significant improvement in the quality of the images reconstructed by deep learning based super-resolution (SR) techniques. In this work, we propose a robust loss function based on the preservation of edges obtained by the Canny operator. This loss function, when combined with the existing loss function such as mean square error (MSE), gives better SR reconstruction measured in terms of PSNR and SSIM. Our proposed loss function guarantees improved performance on any existing algorithm using MSE loss function, without any increase in the computational complexity during testing.

Ram Krishna Pandey, Samarjit Karmakar, A G Ramakrishnan

In this work, we have investigated various style transfer approaches and (i) examined how the stylized reconstruction changes with the change of loss function and (ii) provided a computationally efficient solution for the same. We have used elegant techniques like depth-wise separable convolution in place of convolution and nearest neighbor interpolation in place of transposed convolution. Further, we have also added multiple interpolations in place of transposed convolution. The results obtained are perceptually similar in quality, while being computationally very efficient. The decrease in the computational complexity of our architecture is validated by the decrease in the testing time by 26.1%, 39.1%, and 57.1%, respectively.

A Madhavaraj, T V Ananthapadmanabha, A G Ramakrishnan

Subjective and objective experiments are conducted to understand the extent to which a speaker's gender influences the acoustics of unvoiced (U) sounds. U segments of utterances are replaced by the corresponding segments of a speaker of opposite gender to prepare modified utterances. Humans are asked to judge if the modified utterance is spoken by one or two speakers. The experiments show that human subjects are unable to distinguish the modified from the original. Thus, listeners are able to identify the U segments irrespective of the gender, which may be based on some speaker-independent invariant acoustic cues. To test if this finding is purely a perceptual phenomenon, objective experiments are also conducted. Gender specific HMM based phoneme recognition systems are trained using the TIMIT training set and tested on (a) utterances spoken by the same gender (b) utterances spoken by the opposite gender and (c) the modified utterances of the test set. As expected, the performance is the highest for case (a) and the lowest for case (b). The performance degrades only slightly for case (c). This result shows that the speaker's gender does not as strongly influence the acoustics of U sounds as they do the voiced sounds.

Nazreen P M, A G Ramakrishnan

In this work, we propose the use of dropouts as a Bayesian estimator for increasing the generalizability of a deep neural network (DNN) for speech enhancement. By using Monte Carlo (MC) dropout, we show that the DNN performs better enhancement in unseen noise and SNR conditions. The DNN is trained on speech corrupted with Factory2, M109, Babble, Leopard and Volvo noises at SNRs of 0, 5 and 10 dB and tested on speech with white, pink and factory1 noises. Speech samples are obtained from the TIMIT database and noises from NOISEX-92. In another experiment, we train five DNN models separately on speech corrupted with Factory2, M109, Babble, Leopard and Volvo noises, at 0, 5 and 10 dB SNRs. The model precision (estimated using MC dropout) is used as a proxy for squared error to dynamically select the best of the DNN models based on their performance on each frame of test data.

Ram Krishna Pandey, A G Ramakrishnan

We propose a novel architecture that learns an end-to-end mapping function to improve the spatial resolution of the input natural images. The model is unique in forming a nonlinear combination of three traditional interpolation techniques using the convolutional neural network. Another proposed architecture uses a skip connection with nearest neighbor interpolation, achieving almost similar results. The architectures have been carefully designed to ensure that the reconstructed images lie precisely in the manifold of high-resolution images, thereby preserving the high-frequency components with fine details. We have compared with the state of the art and recent deep learning based natural image super-resolution techniques and found that our methods are able to preserve the sharp details in the image, while also obtaining comparable or better PSNR than them. Since our methods use only traditional interpolations and a shallow CNN with less number of smaller filters, the computational cost is kept low. We have reported the results of two proposed architectures on five standard datasets, for an upscale factor of 2. Our methods generalize well in most cases, which is evident from the better results obtained with increasingly complex datasets. For 4-times upscaling, we have designed similar architectures for comparing with other methods.

Ram Krishna Pandey, Aswin Vasan, A G Ramakrishnan

We propose a computationally efficient architecture that learns to segment lesions from CT images of the liver. The proposed architecture uses bilinear interpolation with sub-pixel convolution at the last layer to upscale the course feature in bottle neck architecture. Since bilinear interpolation and sub-pixel convolution do not have any learnable parameter, our overall model is faster and occupies less memory footprint than the traditional U-net. We evaluate our proposed architecture on the highly competitive dataset of 2017 Liver Tumor Segmentation (LiTS) Challenge. Our method achieves competitive results while reducing the number of learnable parameters roughly by a factor of 13.8 compared to the original UNet model.

Ram Krishna Pandey, A G Ramakrishnan

Recognition of document images have important applications in restoring old and classical texts. The problem involves quality improvement before passing it to a properly trained OCR to get accurate recognition of the text. The image enhancement and quality improvement constitute important steps as subsequent recognition depends upon the quality of the input image. There are scenarios when high resolution images are not available and our experiments show that the OCR accuracy reduces significantly with decrease in the spatial resolution of document images. Thus the only option is to improve the resolution of such document images. The goal is to construct a high resolution image, given a single low resolution binary image, which constitutes the problem of single image super-resolution. Most of the previous work in super-resolution deal with natural images which have more information-content than the document images. Here, we use Convolution Neural Network to learn the mapping between low and the corresponding high resolution images. We experiment with different number of layers, parameter settings and non-linear functions to build a fast end-to-end framework for document image super-resolution. Our proposed model shows a very good PSNR improvement of about 4 dB on 75 dpi Tamil images, resulting in a 3 % improvement of word level accuracy by the OCR. It takes less time than the recent sparse based natural image super-resolution technique, making it useful for real-time document recognition applications.

K V Vijay Girish, A G Ramakrishnan, T V Ananthapadmanabha

A judicious combination of dictionary learning methods, block sparsity and source recovery algorithm are used in a hierarchical manner to identify the noises and the speakers from a noisy conversation between two people. Conversations are simulated using speech from two speakers, each with a different background noise, with varied SNR values, down to -10 dB. Ten each of randomly chosen male and female speakers from the TIMIT database and all the noise sources from the NOISEX database are used for the simulations. For speaker identification, the relative value of weights recovered is used to select an appropriately small subset of the test data, assumed to contain speech. This novel choice of using varied amounts of test data results in an improvement in the speaker recognition rate of around 15% at SNR of 0 dB. Speech and noise are separated using dictionaries of the estimated speaker and noise, and an improvement of signal to distortion ratios of up to 10% is achieved at SNR of 0 dB. K-medoid and cosine similarity based dictionary learning methods lead to better recognition of the background noise and the speaker. Experiments are also conducted on cases, where either the background noise or the speaker is outside the set of trained dictionaries. In such cases, adaptive dictionary learning leads to performance comparable to the other case of complete dictionaries.

K V Vijay Girish, T V Ananthapadmanabha, A G Ramakrishnan

A dictionary learning based audio source classification algorithm is proposed to classify a sample audio signal as one amongst a finite set of different audio sources. Cosine similarity measure is used to select the atoms during dictionary learning. Based on three objective measures proposed, namely, signal to distortion ratio (SDR), the number of non-zero weights and the sum of weights, a frame-wise source classification accuracy of 98.2% is obtained for twelve different sources. Cent percent accuracy has been obtained using moving SDR accumulated over six successive frames for ten of the audio sources tested, while the two other sources require accumulation of 10 and 14 frames.

T V Ananthapadmanabha, K V Vijay Girish, A G Ramakrishnan

Detection of transitions between broad phonetic classes in a speech signal is an important problem which has applications such as landmark detection and segmentation. The proposed hierarchical method detects silence to non-silence transitions, high amplitude (mostly sonorants) to low ampli- tude (mostly fricatives/affricates/stop bursts) transitions and vice-versa. A subset of the extremum (minimum or maximum) samples between every pair of successive zero-crossings is selected above a second pass threshold, from each bandpass filtered speech signal frame. Relative to the mid-point (reference) of a frame, locations of the first and the last extrema lie on either side, if the speech signal belongs to a homogeneous segment; else, both these locations lie on the left or the right side of the reference, indicating a transition frame. When tested on the entire TIMIT database, of the transitions detected, 93.6% are within a tolerance of 20 ms from the hand labeled boundaries. Sonorant, unvoiced non-sonorant and silence classes and their respective onsets are detected with an accuracy of about 83.5% for the same tolerance. The results are as good as, and in some respects better than the state-of-the-art methods for similar tasks.

Deepak Kumar, A G Ramakrishnan

Particle swarm optimization is used in several combinatorial optimization problems. In this work, particle swarms are used to solve quadratic programming problems with quadratic constraints. The approach of particle swarms is an example for interior point methods in optimization as an iterative technique. This approach is novel and deals with classification problems without the use of a traditional classifier. Our method determines the optimal hyperplane or classification boundary for a data set. In a binary classification problem, we constrain each class as a cluster, which is enclosed by an ellipsoid. The estimation of the optimal hyperplane between the two clusters is posed as a quadratically constrained quadratic problem. The optimization problem is solved in distributed format using modified particle swarms. Our method has the advantage of using the direction towards optimal solution rather than searching the entire feasible region. Our results on the Iris, Pima, Wine, and Thyroid datasets show that the proposed method works better than a neural network and the performance is close to that of SVM.

Deepak Kumar, M N Anil Prasad, A G Ramakrishnan

We have benchmarked the maximum obtainable recognition accuracy on various word image datasets using manual segmentation and a currently available commercial OCR. We have developed a Matlab program, with graphical user interface, for semi-automated pixel level segmentation of word images. We discuss the advantages of pixel level annotation. We have covered five databases adding up to over 3600 word images. These word images have been cropped from camera captured scene, born-digital and street view images. We recognize the segmented word image using the trial version of Nuance Omnipage OCR. We also discuss, how the degradations introduced during acquisition or inaccuracies introduced during creation of word images affect the recognition of the word present in the image. Word images for different kinds of degradations and correction for slant and curvy nature of words are also discussed. The word recognition rates obtained on ICDAR 2003, Sign evaluation, Street view, Born-digital and ICDAR 2011 datasets are 83.9%, 89.3%, 79.6%, 88.5% and 86.7% respectively.