Claudius Gläser

Combining Auditory Preprocessing and Bayesian Estimation for Robust Formant Tracking

We present a framework for estimating formant trajectories. Its focus is to achieve high robustness in noisy environments. Our approach combines a preprocessing based on functional principles of the human auditory system and a probabilistic tracking scheme. For enhancing the formant structure in spectrograms we use a Gammatone filterbank, a spectral preemphasis, as well as a spectral filtering using difference-of-Gaussians (DoG) operators. Finally, a contrast enhancement mimicking a competition between filter responses is applied. The probabilistic tracking scheme adopts the mixture modeling technique for estimating the joint distribution of formants. In conjunction with an algorithm for adaptive frequency range segmentation as well as Bayesian smoothing an efficient framework for estimating formant trajectories is derived. Comprehensive evaluations of our method on the VTR-formant database emphasize its high precision and robustness. We obtained superior performance compared to existing approaches for clean as well as echoic noisy speech. Finally, an implementation of the framework within the scope of an online system using instantaneous feature-based resynthesis demonstrates its applicability to real-world scenarios.

Listen to the parrot: Demonstrating the quality of online pitch and formant extraction via feature-based resynthesis

We present a system for online extraction of the fundamental frequency and the first four formant frequencies from a speech signal. In order to evaluate the performance of the extraction a resynthesis of the speech signal is performed. The resynthesis is based on the extracted frequencies and the energy of the input signal at the formant locations. The extraction of the fundamental frequency and the formants is robust against room echoes and interfering noise. In order to improve the robustness against background noise a noise reduction was implemented. Tests in three rooms of different size at varying distances to the system (up to 8 m yielding an SNR of approx. 0 dB) were performed.

Firing Rate Homeostasis for Dynamic Neural Field Formation

Dynamic neural fields are recurrent neural networks which aim at modeling cortical activity evolution both in space and time. A self-organized formation of these fields has been rarely explored previously. The main reason for this is that learning-induced changes in effective connectivity constitute a severe problem with respect to network stability. In this paper, we present a novel network model which is able to self-organize even in face of experience-driven changes in the synaptic strengths of all connections. Key to the model is the incorporation of homeostatic mechanisms which explicitly address network stability. These mechanisms regulate activity of individual neurons in a similar manner as cortical activity is controlled. Namely, our model implements the homeostatic principles of synaptic scaling and intrinsic plasticity. By using fully plastic within-field connections our model further decouples learning from topological constraints. For this reason, we propose to incorporate an additional process which facilitates the development of topology preserving mappings. This process minimizes the wiring length between neurons. We thoroughly evaluated the model using artificial data as well as continuous speech. Our results demonstrate that the network is able to self-organize, maintains stable activity levels, and remains adaptive to variations in input strength and input distribution.

Joint Estimation of Formant Trajectories via Spectro-Temporal Smoothing and Bayesian Techniques

We propose a method for the joint estimation of formant trajectories from spectrograms. Formants are enhanced in the spectrograms obtained from the application of a Gammatone filterbank via a smoothing along the frequency axis. In contrast to previously published approaches, the used tracking algorithm relies on the joint distribution of formants rather than using independent tracker instances. More precisely, Bayesian mixture filtering in conjunction with adaptive frequency range segmentation as well as Bayesian smoothing are used. The algorithm was evaluated on a publicly available database containing hand-labeled formant tracks. Experimental results show a significant performance improvement compared to a state of the art approach.

Learning and use of sensorimotor schemata maps

In this paper we present a framework for the learning and use of sensorimotor schemata. Therefore, we introduce the concept of a schema as a compact representation of an attractor dynamic and discuss how schemata, if embedded into the proposed architecture, can be used to produce, simulate, or recognize goal-directed behaviors. We further present a first implementation of the framework which incorporates well-founded biological principles. Firstly, we apply population coding for the representation of schemata in a neural map and, secondly, we use basis functions as flexible intermediate representations for sensorimotor transformations. Simulation results show that during an initial motor babbling phase the system is able to autonomously develop schemata which correspond to generic behaviors. Moreover, the learned sensorimotor schemata map is topologically ordered insofar as neighboring schemata represent similar behaviors. In accordance with biological findings on the motor system of vertebrates the schemata form a set of behavior primitives which can be flexibly combined to yield more complex behaviors.

Homeostatic development of dynamic neural fields

Dynamic neural field theory has become a popular technique for modeling the spatio-temporal evolution of activity within the cortex. When using neural fields the right balance between excitation and inhibition within the field is crucial for a stable operation. Finding this balance is a severe problem, particularly in face of experience-driven changes of synaptic strengths. Homeostatic plasticity, where the objective function for each unit is to reach some target firing rate, seems to counteract this problem. Here we present a recurrent neural network model composed of excitatory and inhibitory units which can self-organize via a learning regime incorporating Hebbian plasticity, homeostatic synaptic scaling, and self-regulatory changes in the intrinsic excitability of neurons. Furthermore, we do not define a neural field topology by a fixed lateral connectivity; rather we learn lateral connections as well.

An adaptive Normalized Gaussian Network and its application to online category learning

In online applications, where training samples sequentially arise during execution, incremental learning schemes have to be applied. In this paper we propose an adaptive Normalized Gaussian Network model (NGnet) suitable for incremental learning. Following a statistical account we present a truly sequential training procedure. Key to the learning algorithm are local unit manipulation mechanisms for network growth and pruning which continuously adapt the networks complexity according to task demands. We evaluate our model in artificial and real-world categorization tasks. Thereby, we additionally introduce a framework for the categorization on adaptive feature spaces. In the system, a simultaneous extraction of class-discriminative features facilitates the NGnets categorization of input patterns. We present simulation results which demonstrate that the framework realizes a rapid learning from few examples, small-sized network models, and an improved generalization ability. A comparison to incremental support vector machine classification yields a favorable performance of our model.

A computational model for grounding words in the perception of agents

In this paper we present a computational model for incremental word meaning acquisition. It is designed to rapidly build category representations which correspond to the meaning of words. In contrast to existing approaches, our model further extracts word meaning-relevant features using a statistical learning technique. Both category learning and feature extraction are performed simultaneously. To achieve the contradictory needs of rapid as well as statistical learning, we employ mechanisms inspired by Complementary Learning Systems theory. Therefore, our framework is composed of two recurrently coupled components: (1) An adaptive Normalized Gaussian network performs a one-shot memorization of new word-scene associations and uses the acquired knowledge to categorize novel situations. The network further reactivates memorized associations based on its internal representation. (2) Based on the reactivated patterns an additional component subsequently extracts features which facilitate the categorization task. An iterative application of the learning mechanism results in a gradual memory consolidation which let the internal representation of a word meaning become more efficient and robust. We present simulation results for a scenario in which words for object relations concerning position, size, and color have been trained. The results demonstrate that the model learns from few training exemplars and correctly extracts word meaning-relevant features.

Enhancing Topology Preservation during Neural Field Development Via Wiring Length Minimization

We recently proposed a recurrent neural network model for the development of dynamic neural fields [1]. The learning regime incorporates homeostatic processes, such that the network is able to self-organize and maintain a stable operation mode even in face of experience-driven changes in synaptic strengths. However, the learned mappings do not necessarily have to be topology preserving. Here we extend our model by incorporating another mechanism which changes the positions of neurons in the output space. This algorithm operates with a purely local objective function of minimizing the wiring length and runs in parallel to the above mentioned learning process. We experimentally show that the incorporation of this additional mechanism leads to a significant decrease in topological defects and further enhances the quality of the learned mappings. Additionally, the proposed algorithm is not limited to our network model; rather it can be applied to any type of self-organizing maps.