Discriminative utterance verification for connected digits recognition

Abstract

Utterance verification represents an important technology in the design of user-friendly speech recognition systems. It involves the recognition of keyword strings and the rejection of nonkeyword strings. This paper describes a hidden Markov model-based (HMM-based) utterance verification system using the framework of statistical hypothesis testing. The two major issues on how to design keyword and string scoring criteria are addressed. For keyword verification, different alternative hypotheses are proposed based on the scores of antikeyword models and a general acoustic filler model. For string verification, different measures are proposed with the objective of detecting nonvocabulary word strings and possibly erroneous strings (so-called putative errors). This paper also motivates the need for discriminative hypothesis testing in verification. One such approach based on minimum classification error training is investigated in details. When the proposed verification technique was integrated into a state-of-the-art connected digit recognition system, the string error rate for valid digit strings was found to decrease by 57% when setting the rejection rate to 5%. Furthermore, the system was able to correctly reject over 99.9% of nonvocabulary word strings.

Minimization of Utterance Verification Error Rate as a Constrained Optimization Problem

“…In [7], we proposed the MVER training method that directly minimizes the verification error rates 1 . Let N c and N i be the number of correct and incorrect samples in the training data set respectively.…”

“…In general, UV is treated as hypothesis testing [1], [2], [3] using the (log) likelihood ratio test: the ratio between the null hypothesis that the required word is spoken and the alternative hypothesis that it is not. A decision is made by comparing the ratio against a pre-set threshold.…”

String-based minimum verification error (SB-MVE) training for speech recognition

“…Chigier, 1992;Bourlard et al, 1994). In Sukkar (1994), Rahim, Lee and Juang (1995b) and Rose et al (1995) discriminative training methods based on minimum classification error training (Juang & Katagiri, 1992;Chou, Juang, Lee & Soong, 1994a) have been proposed for utterance verification. For example, Rose et al (1995) described a minimum classification error (MCE) training approach for keyword verification which adjusts the parameters of the null hypothesis and the alternative hypothesis models of a tied-mixture density HMM-based system.…”

A discriminative HMM/N-gram-based retrieval approach for mandarin spoken documents

“…However, for simplicity, only the irrelevant document with the highest relevance score, or K=1, is selected for training in this study. There were previous detailed discussions of this issue [Rahim et al 1997;Juang et al 1997]. The classification error function in equation (10) can be transformed into a loss function ranging from 0 to 1 with the Sigmoid operator:…”