Search

Speech signal consists of events in time and frequency, and therefore its analysis with high-resolution time-frequency tools is often of importance. Analytic filter bank provides a simple, fast, and flexible method to construct time-frequency representations of signals. Its parameters can be easily adapted to different situations from uniform to any auditory frequency scale, or even to a focused resolution. Since the Hilbert magnitude values of the channels are obtained at every sample, it provides a practical tool for a high-resolution time-frequency analysis. The present study describes the basic theory of analytic filters and tests their main properties. Applications of analytic filter bank to different speech analysis tasks including pitch period estimation and pitch synchronous analysis of formant frequencies and bandwidths are demonstrated. In addition, a new feature vector called group delay vector is introduced. It is shown that this representation provides comparable, or even better results, than those obtained by spectral magnitude feature vectors in the analysis and classification of vowels. The implications of this observation are discussed also from the speech perception point of view.

Automatic detection and classification of short and nonstationary events in noisy signals is widely considered to be a difficult task for traditional frequency domain and even time–frequency domain approaches. A novel method for audio signal classification is introduced. It is based on statistical properties of the temporal fine structure of audio events. Artificially generated random signals and unvoiced stop consonants of speech are used to evaluate the method. The results show improved recognition accuracy in comparison to traditional approaches.

An efficient method for weakly supervised pattern discovery and recognition from discrete categorical sequences is introduced. The method utilizes two parallel sources of data: categorical sequences carrying some temporal or spatial information and a set of labeled, but not exactly aligned, contextual events related to the sequences. From these inputs the method builds associative models able to describe systematically co-occurring structures in the input streams. The learned models, based on transitional probabilities of events observed at several different time lags, inherently segment and classify novel sequences into contextual categories. Learning and recognition processes are purely incremental and computationally cheap, making the approach suitable for on-line learning tasks. The capabilities of the algorithm are demonstrated in a keyword learning task from continuous infant-directed speech and a continuous speech recognition task operating at varying noise levels.

Identification of bird species based on their vocalization is studied in this paper. The main focus is introducing a new parametric representation of bird sounds for automatic identification of their species. The method is based on the statistics of local temporal patterns in bird vocalization. Two different sets of bird species are used in the classification tests. The first set contains six species that often produce inharmonic sounds. For the second set, four species that produce very different types of sounds were added. Recognition results using a k-NN-classifier shows improved recognition accuracy over the results obtained by MFCC-features.

Since modern computational devices are required to store and process increasing amounts of data generated from various sources, efficient algorithms for identification of significant information in the data are becoming essential. Sensory recordings are one example where automatic and continuous storing and processing of large amounts of data is needed. Therefore, algorithms that can alleviate the computational load of the devices and reduce their storage requirements by removing uninformative data are important. In this work we propose a method for data reduction based on theories of human attention. The method detects temporally salient events based on the context in which they occur and retains only those sections of the input signal. The algorithm is tested as a pre-processing stage in a weakly supervised keyword learning experiment where it is shown to significantly improve the quality of the codebooks used in the pattern discovery process.

A novel and computationally straightforward clustering algorithm was developed for vector quantization (VQ) of speech signals for a task of unsupervised pattern discovery (PD) from speech. The algorithm works in purely incremental mode, is computationally extremely feasible, and achieves comparable classification quality with the well-known k-means algorithm in the PD task. In addition to presenting the algorithm, general findings regarding the relationship between the amounts of training material, convergence of the clustering algorithm, and the ultimate quality of VQ codebooks are discussed. Index Terms: speech recognition, pattern discovery, time series analysis, vector quantization, data clustering

An efficient method for pattern discovery from discrete time series is introduced in this paper. The method utilizes two parallel streams of data, a discrete unit time-series and a set of labeled events, From these inputs it builds associative models between systematically co-occurring structures existing in both streams. The models are based on transitional probabilities of events at several different time scales. Learning and recognition processes are incremental, making the approach suitable for online learning tasks. The capabilities of the algorithm are demonstrated in a continuous speech recognition task operating in varying noise levels.

A vector autoregressive (VAR) model is used in the auditorytime-frequency domain to predict spectral changes. Forwardand backward prediction errors increases at the phone bound-aries. These error signals are then used to study and detect theboundaries of the largest changes allowing the most reliableautomatic segmentation. Using a fully unsupervised methodyields segments consisting of a variable number of phones. Thequality of performance of this method was tested with a set of150 Finnish sentences pronounced by one female and two malespeakers. The performance for English was tested using theTIMIT core test set. The boundaries between stops and vowels,in particular, are detected with high probability and precision. 1. Introduction Many subfields of speech technology need robust methodsfor automatic phonetic speech segmentation. Preferably thesemethods would be fully speaker and language independent.They should perform segmentation without any prior infor-mation about the speaker or the utterance in question. Thesemethods should not apply any type of prior learning, andthey should be able to process unknown utterances in a fullyunsupervised manner. This paper describes a preliminary testof a novel method for automatic speech segmentation, whichfulfills the hard demands mentioned to a certain degree.Segmentation methods described in the literature can beclassified into explicit and implicit methods. They also varyin terms of segmentation units (e.g. phonemes, syllables,words). In explicit methods, the underlying phoneme sequenceis known prior to the segmentation. These methods are usedin speech synthesis for example. Implicit methods split theutterance into smaller units without using any informationabout the underlying phoneme sequence. These methods arebased on analyzing the acoustic properties of the signal anddetecting either spectrally stable parts or rapid variations ofsignal. An example of a method based on locating spectrallystable parts is in [1] where the correlation between parameterscomputed from nearby frames has been used as a measure ofstability. In [2], segment boundaries are implicitly detectedcomparing the means of frames around potential boundariesusing “jump-function.” In [3], the variations of short-termenergy function is used as a measure to produce syllable-likeunits using minimum phase group delay functions.In the case of continuous speech, the signal cannot bestrictly divided into stable and varying parts which wouldcorrespond one-to-one with phones and segment boundaries.No phone in continuous speech produces steady spectra,but instead within a phone there are always slow spectralmovements which are, to some degree, possible to predict.The method proposed in this paper does not detect these slowspectral variations, but rather is based on detecting unpredictedchanges in auditory time-frequency picture of speech at phoneboundaries. These unpredicted changes happen most oftenwhen moving from one phoneme class to another. Change inthe speech production mechanism changes the acoustic signalin an unpredictable manner. Knowing that not all transitionsproduce a large or rapid spectral change, a question of thisstudy is which kind of phone boundaries allows the mostreliable and robust detection by the method.When facing speaker-independent unlimited vocabulary(e.g. inflectional languages) continuous speech recognition, thewords have to be split into smaller units such as morphemes;hence, not every phone boundary needs to be detected. Seg-ments similar to syllables or morphemes consisting of one tomany phones do apply as well as long as the total number ofdifferent segments is not too high for modeling purposes.The novel method presented in this paper produces seg-ments consisting of phone clusters of different lengths. Thecore idea is to model the spectral variation by using VectorAutoregressive model (VAR). The model performs forward andbackward predictions in the auditory time-frequency domainwith associated prediction errors. The segment boundarycandidates are found based on these error signals.