Publications

In this paper, we present a conditional pronunciation modeling method for the speaker detection task that does not rely on acoustic vectors. Aiming at exploiting higher-level information carried by the speech signal, it uses time-aligned streams of phones and phonemes to model a speaker's specific Pronunciation. Our system uses phonemes drawn from a lexicon of pronunciations of words recognized by an automatic speech recognition system to generate the phoneme stream and an open-loop phone recognizer to generate a phone stream. The phoneme and phone streams are aligned at the frame level and conditional probabilities of a phone, given a phoneme, are estimated using co-occurrence counts. A likelihood detector is then applied to these probabilities. Performance is measured using the NIST Extended Data paradigm and the Switchboard-I corpus. Using 8 training conversations for enrollment, a 2.1% equal error rate was achieved. Extensions and alternatives, as well as fusion experiments, are presented and discussed.

The area of automatic speaker recognition has been dominated by systems using only short-term, low-level acoustic information, such as cepstral features. While these systems have indeed produced very low error rates, they ignore other levels of information beyond low-level acoustics that convey speaker information. Recently published work has shown examples that such high-level information can be used successfully in automatic speaker recognition systems and has the potential to improve accuracy and add robustness. For the 2002 JHU CLSP summer workshop, the SuperSID project was undertaken to exploit these high-level information sources and dramatically increase speaker recognition accuracy on a defined NIST evaluation corpus and task. This paper provides an overview of the structures, data, task, tools, and accomplishments of this project. Wide ranging approaches using pronunciation models, prosodic dynamics, pitch and duration features, phone streams, and conversational interactions were explored and developed. In this paper we show how these novel features and classifiers indeed provide complementary information and can be fused together to drive down the equal error rate on the 2001 NIS extended data task to 0.2% - a 71% relative reduction in error over the previous state of the art.

While there has been a long tradition of research seeking to use prosodic features, especially pitch, in speaker recognition systems, results have generally been disappointing when such features are used in isolation and only modest improvements have been set when used in conjunction with traditional cepstral GMM systems. In contrast, we report here on work from the JHU 2002 Summer Workshop exploring a range of prosodic features, using as testbed NIST's 2001 Extended Data task. We examined a variety of modeling techniques, such as n-gram models of turn-level prosodic features and simple vectors of summary statistics per conversation side scored by kth nearest-neighbor classifiers. We found that purely prosodic models were able to achieve equal error rates of under 10%, and yielded significant gains when combined with more traditional systems. We also report on exploratory work on "conversational" features, capturing properties of the interaction across conversion sides, such as turn-taking patterns.