Search

Gender and age estimation based on Gaussian Mixture Models (GMM) is introduced. Telephone recordings from the Czech SpeechDat-East database are used as training and test data set. Mel-Frequency Cepstral Coefficients (MFCC) are extracted from the speech recordings. To estimate the GMMs' parameters Maximum Likelihood (ML) training is applied. Consequently these estimations are used as the baseline for Maximum Mutual Information (MMI) training. Results achieved when employing both ML and MMI training are presented and discussed.

One essential security weakness in many modern systems is the difficulty of managing secured and provable-identities for all participating entities. In this work we introduce a process for generating a secret provable identity for electronic devices. The identity is created through a random process that is triggered as an electronic-Mutation (EM) once at a user-defined time after device manufacture. The result should be a provable, certified, secret, unclonable, and unchangeable identity that can serve as an electronic DNA (e-DNA) for the device. The identity is self- created similar to naturally occurring Physical Unclonable Functions (PUF), such that it is infeasible even for the unit manufacturer to create duplicate identities. The proposed identity should in difference to PUF evolve through the lifetime of the device allowing for easier detection of fraud attacks. The identity should also possibly diffuse in all system entities similar to the diffusion of biological DNA in all entities of a living organism. Many legal and criminal issues plaguing mass public-commerce, e-Government and mobile systems could be easily resolved through such an identification technology.

This paper describes "search in speech" techniques developed in the Speech@FIT research group at FIT BUT in the last couple of years. It concentrates on spoken term detection (STD) and presents our system for NIST STD 2006 evaluations in detail. It also briefly mentions our systems for speaker and language recognition.

In recent years, probabilistic features became an integral part of state-of-the-are LVCSR systems. In this work, we are exploring the possibility of obtaining the features directly from neural net without the necessity of converting output probabilities to features suitable for subsequent GMM-HMM system. We experimented with 5-layer MLP with bottle-neck in the middle layer. After training such a neural net, we used outputs of the bottle-neck as features for GMM-HMM recognition system. The benefits are twofold: first, improvement was gained when these features are used instead of the probabilistic features, second, the size of the system was reduced, as only part of the neural net is used. The experiments were performed on meetings recognition task defined in MST RT'05 evaluation.

This paper describes STBU 2006 speaker recognition system, which performed well in the NIST 2006 speaker recognition evaluation. STBU is consortium of 4 partners: Spescom Data Voice (South Africa), TNO (Netherlands), BUT (Czech Republic) and University of Stellenbosch (South Africa). The primary system is a combination of three main kinds of systems: (1) GMM, with short-time MFCC or PLP features, (2) GMM-SVM, using GMM mean supervectors as input and (3) MLLR-SVM, using MLLR speaker adaptation coefficients derived from English LVCSR system. In this paper, we describe these sub-systems and present results for each system alone and in combination on the NIST Speaker Recognition Evaluation (SRE) 2006 development and evaluation data sets. © 2007 IEEE.

This paper presents comparison of Maximum Likelihood (ML) and discriminative Maximum Mutual Information (MMI) training for acoustic modeling in language identification (LID). Both approaches are compared on state-of- the-art shifted delta-cepstra features, the results are reported on data from NIST 2003 evaluations. Clear advantage of MMI over ML training is shown. Further improvements of acoustic LID are discussed: Heteroscedastic Linear Discriminant Analysis (HLDA) for feature de-correlation and dimensionality reduction and Ergodic Hidden Markov models (EHMM) for better modeling of dynamics in the acoustic space. The final error rate compares favorably to other results published on NIST 2003 data.

This paper presents the language identification (LID) system developed in Speech@FIT group at Brno University of Technology (BUT) for NIST 2005 Language Recognition Evaluation. The system consists of two parts: phonotactic and acoustic. Phonotactic system is based on hybrid phoneme recognizers trained on SpeechDat-E database. Phoneme lattices are used to train and test phonotactic language models. Further improvement is obtained by using anti-models. Acoustic system is based on GMM modeling trained under Maximum Mutual Information framework. We describe both parts and provide a discussion of performance on LRE 2005 recognition task.

The CARETAKER project, which is a 30-month project that has just kicked off, aims at studying, developing and assessing multimedia knowledge-based content analysis, knowledge extraction components, and metadata management sub-systems in the context of automated situation awareness, diagnosis and decision support. More precisely, CARETAKER will focus on the extraction of structured knowledge from large multimedia collections recorded over networks of cameras and microphones deployed in real sites. The produced audio-visual streams, in addition to surveillance and safety issues, could represent a useful source of information if stored and automatically analyzed, in urban/environment planning, resource optimization, disabled/elderly person monitoring

The problem of very low bit rate segmental speech coding is addressed. The basic units are found automatically in the training database using temporal decomposition, vector quantization and multigrams. They are modelled by HMMs. The coding is based on recognition and synthesis. In single speaker tests, we obtained intelligible and naturally sounding speech at a mean rate of 211.2 b/s. In the end, future extensions of our scheme (diphone-like synthesis and speaker adaptation) as well as possible use of automatically derived units in recognition are discussed.

The aim of our effort is to reach higher quality of the resulting speech coded by a very low bit rate (VLBR) segmental coder. The basic units are found automatically in a training database using temporal decomposition and vector quantization. They are modeled by HMM. Then two methods of re-segmentation are used in order to find new longer units. In the first approach borders are set to the centers of previous units. In the second, borders are fixed to the centers of middle HMM states of previous units. The number of frames in new units is conditioned to be bigger than a fixed constant. Hence, new units can consist of several previous segments. Decreasing transition noise of the resultant speech was obtained using these techniques.

Cet article presente les principales techniques de codage de parole a bas et tres bas debits, de 50 bit/s a 4 000 bit/s. Puis il presente en detail la methode hsx pour le codage a 1200 bit/s et une nouvelle approche segmentale utilisant des unites acoustiques obtenues de maniere non supervisee pour des debits inferieurs a 400 bit/s.

International audience

Training deep recurrent neural network (RNN) architectures is complicated due to the increased network complexity. This disrupts the learning of higher order abstracts using deep RNN. In case of feed-forward networks training deep structures is simple and faster while learning long-term temporal information is not possible. In this paper we propose a residual memory neural network (RMN) architecture to model short-time dependencies using deep feed-forward layers having residual and time delayed connections. The residual connection paves way to construct deeper networks by enabling unhindered flow of gradients and the time delay units capture temporal information with shared weights. The number of layers in RMN signifies both the hierarchical processing depth and temporal depth. The computational complexity in training RMN is significantly less when compared to deep recurrent networks. RMN is further extended as bi-directional RMN (BRMN) to capture both past and future information. Experimental analysis is done on AMI corpus to substantiate the capability of RMN in learning long-term information and hierarchical information. Recognition performance of RMN trained with 300 hours of Switchboard corpus is compared with various state-of-the-art LVCSR systems. The results indicate that RMN and BRMN gains 6 % and 3.8 % relative improvement over LSTM and BLSTM networks.

Contains fulltext : 76438.pdf (author's version ) (Open Access) 4 p.

This paper concentrates on PRLM (phoneme recognizer followed by language model) approach to language recognition. It elaborates on our prior work concerning the quality of phoneme recognition and amounts of training data for phoneme recognizer training. It reports improvements brought to our PRLM system by better phoneme recognition and Witten-Bell discounting in LM-modeling. The paper then concentrates on the use of phoneme lattices and anti-models. Training and scoring on phoneme lattices brought significant improvement in language recognition accuracy. The antimodels are simple, yet powerful technique to improve the discrimination between target and non-target languages. All results are reported on standard NIST 2003 data; comparison with other published results is favorable to our system.

Current systems for recognition, synthesis, very low bit-rate (VLBR) coding and text-independent speaker verification rely on sub-word units determined using phonetic knowledge. This paper presents an alternative to this approach - determination of speech units using AUSP (Automatic Language Independent Speech Processing) tools. Experimental results for speaker-dependent VLBR coding are reported on two databases: average rate of 120 bps for unit encoding was achieved. In verification, this approach was tested during 1998's NIST-NSA evaluation campaign with a MLP-based scoring system.

This paper deals with phoneme recognition based on neural networks (NN). First, several approaches to improve the phoneme error rate are suggested and discussed. In the experimental part, we concentrate on TempoRAl Patterns (TRAPs) and novel split temporal context (STC) phoneme recognizers. We also investigate into tandem NN architectures. The results of the final system reported on standard TIMIT database compare favorably to the best published results.