We propose a novel method of measuring the similarity between two or more speech utterances for speaker clustering, based on probability theory and factor analysis. The similarity function is formulated as the probability that the utterances originated from the same speaker, and uses statistical eigenvoice and eigenchannel models to incorporate physical knowledge of interspeaker and intraspeaker variabilities, allowing the similarity function to be trainable and robust. The comparison function can be efficiently computed using a compact set of sufficient statistics for each speech utterance, allowing the acoustic features to be discarded. We begin using only eigenvoices, and then show how the eigenchannels can be incorporated into the equation to result in an identical form but with a different set of sufficient statistics. We test the proposed model in a speaker clustering task using the CALLHOME telephone conversation corpus and show that it performs better than two other well-known similarity measures: the Cross-Likelihood Ratio (CLR) and Generalized Likelihood Ratio (GLR).

We propose a method of music melody matching based on their "continuous" (or "time-frame-based") pitch contours. Most previous methods using frame-based contours either made limiting assumptions on the locations, musical scale, tempo, and/or rhythm of the queries in relations to the targets, or involved exhaustive dynamic time-warping procedures that were too computation-intensive for use in real scenarios. In the proposed method, variable-scale windowing and wavelet transformations are performed at an initial coarse search stage to efficiently match queries to targets. At the following fine search stage, we apply a novel segmented dynamic time warping (DTW) method for melody contours, computing a more accurate distance between the query and each of the candidate targets with less computation than traditional DTW. The method searches arbitrary target locations and explicitly adjusts for differences in tempo and musical scale between queries and targets as well as rhythmic inconsistencies within queries. At the same time, retrieval times are fast enough for use in real-world scenarios.

We propose a novel utterance comparison model based on probability theory and factor analysis that computes the likelihood of two speech utterances originating from die same speaker. The model depends only on a set of statistics extracted from each utterance and can efficiently compare utterances using these statistics without requiring die indefinite storage of speech features. We apply the model as a distance metric for speaker clustering in die CALLHOME telephone conversation corpus to achieve competitive results compared to three other known similarity measures: the Generalized Likelihood Ratio, Cross-Likelihood Ratio, and eigenvoice distance.

In this paper, we present an efficient method of speech indexing and search using phoneme sequences called uniterms. In the indexing stage, a collection of uniterms and uniterm sequences is extracted from the target speech database by applying statistical scoring to each data item's phoneme lattice. In the search stage, each speech query's phoneme lattice is used to select candidate uniterms from the collection. These uniterms are applied in a speech recognition engine to convert the speech query into a uniterm lattice, from which we obtain a set of candidate uniterm sequences, each of which can be mapped to a search result item. Not only is this method a significant improvement over previous phoneme-based methods, it is shown that explicit sequential comparison of uniterms in query and target data can be avoided using the proposed method without loss of search performance. Avoiding sequential comparison allows better handling of transposition of words, and for the case where queries have word orders different from their intended targets, the proposed method can potentially bring about significant improvement.

In this paper we present two novel approaches for voice indexing and search. The first approach is a Uniterm based voice indexing and search scheme that can be used for the fast retrieval of voice tagged multimedia contents on mobile devices. Uniterms, a string of phonemes with high scores, are extracted, in the indexing stage, from the phoneme lattice. For retrieval, the Uniterms are scored against the latent lattice model from the query voice. The candidate Uniterm list is selected to generate the audio segments of which the best phoneme paths are retrieved. These best paths of the candidate audio segments are compared against the best path of query voice and the search results are generated among the best matches. The second approach is also presented for comparison where the indices are extracted from the lattice in the form of unigram and bigram feature vectors. Each audio segment is represented by a tf-idf modulated feature vector. The search process involves two stages: the coarse search looks up the index and quickly returns a set of candidates; the fine search then compares the best paths of the query voice to the phone lattices of the candidates by using dynamic programming. Experimental results show that the Uniterm approach is significantly better and it is feasible for such voice search approaches on voice tagged multimedia contents on mobile devices. Finally we introduce a real-time news broadcasting search system.

To improve the performance and the usability of the speech recognition devices, it is necessary for most applications to allow users to enter new words or personalize words in the system vocabulary. The voicetagging technique is a simple example of using speaker dependent spoken samples to generate baseform transcriptions of the spoken words. More sophisticated techniques can use both spoken samples and text versions of the new words to generate baseform transcriptions. In this paper, we propose a maximum context tree (MCT) based approach to the problem. Comparison is made to the common decision tree based method and Pronunciation by Analogy (PbA) approach. The new approach gives exact baseform transcription for in-vocabulary words and it shows better performance than decision tree. It performs significantly better than PbA approach with less memory usages. MCT uses the word segment probability rather than frequency count used in PbA. MCT uses the full context for the focus letter to overcome the some deficiencies in the PbA approach.

Comparing speech recognition performance between human beings and computer, the latter's performance degrades dramatically in a noisy mobile environment. Based on the perceptual study of speech sounds in terms of auditory masking and speech perception study, it has been suggested that the contribution of each auditory channel is rather independent as used in articulation index model. From the point of view of the signal process perspective, we can diminish the noise caused variability by using spectral masking so that the noise signal between spectrum gaps is masked and limit the error propagation between frequency (auditory) channels. Based on this observation, we propose a novel robust feature extraction based spectrally masked channel energy ratio (SMaChER). We show the significant error reduction rate on noisy data has been achieved.

Support vector machines represent a new approach to pattern classification developed from the theory of structural risk minimization. In this paper, we present an investigation into the application of support vector machines to the confidence measurement problem in speech recognition. Specifically, based on the results from an initial decoding of an utterance during speech recognition, we derive a feature vector consisting of parameters such as word score density, N-best word score density differences, relative word score and relative word duration as input to the confidence measurement process in which hypothetically correct utterances are accepted and utterances determined to be incorrect are rejected. We propose a new approach to training support vector machines. In this paper, we train and test a support vector machines classifier and compare the results with other statistical classification methods.

To improve the performance and the usability of the speech recognition devices, It is necessary for most applications to allow users to enter new words or personalize words to the system vocabulary. Voice-tagging technique is a simple example that use speaker dependent spoken sample to generate baseform transcriptions of the spoken words. More sophisticated techniques can use both spoken samples and texts of the new words to generate baseform transcriptions. In this paper, we propose a new approach to the problem. We use Bayesian networks to model the letter-to-sound rule probabilities. Compared to the common decision tree based method, This new approach shows a definite advantage.

The detection of glottal closure instants has been a necessary step in several applications of speech processing, such as voice source analysis, speech prosody manipulation and speech synthesis. The paper presents a new algorithm for glottal closure detection that compares favorably with other methods available in terms of robustness and computational efficiency. The authors propose to use the singular value decomposition (SVD) approach to detect the instants of glottal closure from the speech signal. The proposed SVD method amounts to calculating the Frobenius norms of signal matrices and therefore is computationally efficient. Moreover, it produces well-defined and reliable peaks that indicate the instants of glottal closure. Finally, with the introduction of the total linear least squares technique, two other proposed methods are reinvestigated and unified into the SVD framework.>

This paper establishes the relationship between a weighted linear prediction method used for robust analysis of voiced speech and the autoregressive modelling based on higher-order statistics, known as cumulants

The detection of narrowband noise targets in maskers of equal-amplitude harmonics has been reported in an earlier contribution [4]. The results revealed that the masking patterns are strongly dependent on the fundamental frequency of the masker. In the present paper we report masking patterns of broadband harmonic complexes as a function of their spectral tilt and level, and the masking pattern of a vowel sound. We will focus on how the spectral masking patterns are influenced by temporal processes in the auditory system