This paper presents a possible application of a text-dependent speaker recognition system within the unconstrained domain of telephone conversation speech, as contained in the Switchboard I corpus. The system utilizes word HMMs to generate likelihood scores for key words among the backchannel, filled pause, and discourse marker categories. Results on tests using a variant of the NIST 2001 exten...

This article presents the RIMO/ELISA speaker veriication system which has been used in the 1999 NIST speaker recognition evaluation. This system is based on a new technique for analyzing speech signals called time-frequency principal component (TFPC) analysis. This technique consists in extracting principal components from the contextual covariance matrix, which is the covariance matrix of a se...

We argue for a surrcial pronunciation model: a model without underlying forms. The surrcial model outper-forms a traditional generative model by a signiicant margin on conversational speech (Switchboard) as well as on read speech (TIMIT). Our results suggest that the true mapping from underlying forms to surface forms is too complex to be accurately modeled using current techniques, and that we...

We identify a set of prosodic cues for parsing conversational speech and show how such features can be effectively incorporated into a statistical parsing model. On the Switchboard corpus of conversational speech, the system achieves improved parse accuracy over a state-of-the-art system which uses only lexical and syntactic features. Since removal of edit regions is known to improve downstream...

We report the results of some experiments which demonstrate that eigenvoice MAP and eigenphone MAP are at least as effective as classical MAP for discriminative speaker modeling on SWITCHBOARD data. We show how eigenvoice MAP can be modified to yield a new model-based channel compensation technique which we call eigenchannel MAP. When compared with multi-channel training, eigenchannel MAP was f...

In this paper text-independent automatic speaker verification based on support vector machines is considered. A generalized linear kernel training method based on kernel alignment maximization is proposed. First, kernel matrix decomposition into a sum of maximally aligned directions in the input space is performed and this decomposition is spectrally optimized. The method was evaluated for high...

We define the task of incremental or 0lag utterance segmentation, that is, the task of segmenting an ongoing speech recognition stream into utterance units, and present first results. We use a combination of hidden event language model, features from an incremental parser, and acoustic / prosodic features to train classifiers on real-world conversational data (from the Switchboard corpus). The ...

This paper proposes a probabilistic framework to de ne and evaluate con dence measures for word recognition. We describe a novel method to combine di erent knowledge sources and estimate the con dence in a word hypothesis, via a neural network. We also propose a measure of the joint performance of the recognition and con dence systems. The de nitions and algorithms are illustrated with results ...

We present a novel approach to spoken dialogue summarization. Our system employs a set of semantic similarity metrics using the noun portion of WordNet as a knowledge source. So far, the noun senses have been disambiguated manually. The algorithm aims to extract utterances carrying the essential content of dialogues. We evaluate the system on 20 Switchboard dialogues. The results show that our ...

In what dialog situations and contexts do backchannels commonly occur? This paper examines this question using a newly developed notion of dialog space, defined by orthogonal, prosody-derived dimensions. Taking 3363 instances of uh-huh, found in the Switchboard corpus, we examine where in this space they tend to occur. While the results largely agree with previous descriptions and observations,...