Patricia A. Pelle

A Robust Pitch Extraction System Based on Phase Locked Loops

We propose a new scheme for pitch estimation using phase locked loops (PLL). These devices possess desired properties for speech signal analysis, including instantaneous phase detection, and noise robustness. The method is based on a PLL arrange applied to the outputs of a bank of passband filters, each of them tuned to a portion of the spectrum of the signal. PLL devices provide us with robust information about the period of the speech signal harmonics. Combining this information with the spectrum of selected portions of the signal, it is possible to determine a reliable estimate of the period of the whole signal. Performance is evaluated by comparing our system to get_f0 algorithm, with various noise levels. We show that our system largely outperforms get_f0 under noisy speech conditions

Robust speech representation of voiced sounds based on synchrony determination with PLLs

We propose to include synchrony effects, known to exist in the auditory system, to represent voiced parts of the speech signal in a robust way. The system decomposes the input signal by means of a bandpass filter bank, and utilizes a bank of phase locked loops (PLLs) to obtain information on the frequencies present at a specific time. This information about the frequency distribution is transformed into a spectral-like representation based on synchrony effects. Noisy speech recognition experiments are performed using this synchrony-based spectrum, which is transformed into a small set of coefficients by using a transformation similar to that utilized for mel cepstrum features. We show that recognition performance compared to mel cepstrum features is advantageous, when measured over a range of SNR conditions, especially in the high noise level case.

Analysis of the BSB Model Dynamics Using Control Theory

A novel method to analyze the dynamics of the BSB (Brain State in a Box) model is presented. The method is able to determine if a proposed interconnection matrix generated with the CMM rule can achieve the desired behavior, and what are the parameters that should be changed in case of obtaining an incorrect final rest point. By means of an application of the techniques used in control system theory it is possible to evaluate the evolution of each of the training points individually and also to establish what are the points of the training set that are competing between them. So, this kind of analysis can be useful not only as a medium to modify the learning rule to obtain a better performance, but also as a tool to understand the limitations that the initial learning rule has. As an example of the application of this tool some simulated experiments were carried out in order to show the effectiveness of the proposed method.

Speaker identification system based in verification techniques with Bayesian discrimination

In this work an automatic speaker identification system of good performance is presented, which is based in state-of-the-art verification systems. The problem of speaker verification can be extended to the task of identifying an individual between a group, called speaker identification. In this paper a modification to the final Bayesian discrimination between different speakers is proposed, including classification within the theoretical setting of the final statistical model. The system is able to display the usual good verification systems results, even for restricted data conditions.

Keyword detection system with zero-resources techniques

A keyword detection system with zero-resources techniques is presented. It consists of a primary alignment method and a later rescoring of its hipotheses. Both stages based on a segmental dynamic time warping method and a segmental model respectively. The resulting system is totally language independent and has no pre-segmentation or written resources requirements, being its only needed input information a set of stored keyword emissions. Recognition rates of 60% with 15% of false positive errors are obtained, and is clearly demonstrated that the rescoring stage improves overall system performance.

A robust pitch detector based on time envelope and individual harmonic information using Phase Locked Loops and consensual decisions

A pitch detector system is proposed, based on estimation of time envelope information and individual harmonics frequency determination. The system takes advantage of the fact that both the time envelope of a speech signal and the frequency of its individual harmonics carry information about the pitch. A set of pitch hypotheses are calculated from time envelope signals extracted from different parts of the spectrum. These hypotheses are tested against the information obtained from individual harmonics from the lower part of the spectrum. Hypotheses which reliably match that information are preserved, and a consensual decision is taken between them to obtain the final pitch estimation. In addition, several by-products of the process can be retained in order to perform voiced/unvoiced detection. As every subsystem extracts information from arrangements of Phase Locked Loops the detector is very robust in noisy conditions, compared to the well known get f0 algorithm.