A two-level classifier for text-independent speaker identification

Hadjitodorov, Stefan; Boyanov, B.; Dalakchieva, N.

doi:10.1016/s0167-6393(97)00004-6

Cited by 9 publications

(3 citation statements)

References 7 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…With no or very limited a priori knowledge about the processes studied, the empirical ANN model compensates for its inherent information inadequacy by requiring fairly large and well-spread training sets (38). Once a good training set of input-;output data is available, however, ANN models can prove useful for specific applications.…”

Section: Resultsmentioning

confidence: 99%

Empirical versus mechanistic modelling: Comparison of an artificial neural network to a mechanistically based model for quantitative structure pharmacokinetic relationships of a homologous series of barbiturates

et al. 1999

View full text Add to dashboard Cite

The aim of the current study was to compare the predictive performance of a mechanistically based model and an empirical artificial neural network (ANN) model to describe the relationship between the tissue-to-unbound plasma concentration ratios (Kpu's) of 14 rat tissues and the lipophilicity (LogP) of a series of nine 5-nalkyl-5-ethyl barbituric acids. The mechanistic model comprised the water content, binding capacity, number of the binding sites, and binding association constant of each tissue. A backpropagation ANN with 2 hidden layers (33 neurons in the first layer, 9 neurons in the second) was used for the comparison. The network was trained by an algorithm with adaptive momentum and learning rate, programmed using the ANN Toolbox of MATLAB. The predictive performance of both models was evaluated using a leave-one-out procedure and computation of both the mean prediction error (ME, showing the prediction bias) and the mean squared prediction error (MSE, showing the prediction accuracy). The ME of the mechanistic model was 18% (range, 20 to 57%), indicating a tendency for overprediction; the MSE is 32% (range, 6 to 104%). The ANN had almost no bias: the ME was 2% (range, 36 to 64%) and had greater precision than the mechanistic model, MSE 18% (range, 4 to 70%). Generally, neither model appeared to be a significantly better predictor of the Kpu's in the rat.

show abstract

Section: Resultsmentioning

confidence: 99%

Empirical versus mechanistic modelling: Comparison of an artificial neural network to a mechanistically based model for quantitative structure pharmacokinetic relationships of a homologous series of barbiturates

et al. 1999

View full text Add to dashboard Cite

show abstract

“…A two level classifier for closed‐set speaker identification was presented by Hadjitodorov et al . (). The paper investigates different versions of a SOM as a first stage classifier to obtain a prototype distribution map and then use these maps to feed a second‐stage MLP network classifier.…”

Section: Related Workmentioning

confidence: 97%

Speaker verification using heterogeneous neural network architecture with linear correlation speech activity detection

2013

View full text Add to dashboard Cite

This paper presents a multi-level speaker verification system that uses 64 discrete Fourier transform spectrum components as input feature vectors. A speech activity detection technique is used as a pre-processing stage to identify vowel phoneme boundaries within a speech sample. A modified self-organising map (SOM) is then used to filter the speech data by using cluster information extracted from three vowels for a claimed speaker. This SOM filtering stage also provides coarse speaker verification. Finally, a second speaker verification level of three multi-layer perceptron networks classifies the filtered frames provided by the SOMs. These multi-layer perceptrons work as fine-grained vowel-based speaker verifiers. The proposed verification algorithm shows a performance of 94.54% when evaluated using 50 speakers from the Centre for Spoken Language Understanding speaker verification database. In addition, it is shown that the novel discrete Fourier transform spectrum-based linear correlation pre-processing technique, presented here, provides the system with greater robustness against changes in speech volume levels when compared with an equivalent energy frame analysis.

show abstract

“…The artificial neural networks (ANN) that are known as a simplified model of biological nervous system (Katagiri, 2000;Bishop, 1995) allow for such observation of the non-fluent speech analysis. That is why the growing trend in the ANN or ANN and HMMs hybrid applications could be observed in such areas as speech and speaker recognition (Farrell, 2000;Farrell et al, 1994;Hadjitodorov et al, 1997;Leung et al, 2007;Trentin and Giuliani, 2001) and classification (Hosom, 2003;Cosi et al, 2000;Kocsor et al, 2000;Lee et al, 1998) or feature extraction (Katagiri, 2000;Gemello et al, 2007;Fritsch et al, 2000;Yegnanarayana and Narendranath, 2000;Shao and Barker, 2008;Uncini, 2003). The Multilayer Perceptron (MLP) and Radial Basis Function networks (RBF) as well as recurrent and fuzzy networks frequently occur in the automatic speech recognition (ASR) process (Chen et al, 1996;Farrell, 2000;Leung et al, 2007;Schuster, 2000).…”

Section: Introductionmentioning

confidence: 99%