Effective Utterance Classification with Unsupervised Phonotactic Models

Effective Utterance Classification with Unsupervised Phonotactic Models Hiyan Alshawi AT&T Labs - Research Florham Park, NJ 07932, USA hiyan@research.att.com Abstract This paper describes a method for utterance classification that does not require manual transcription of training data. The method combines domain independent acoustic models with off-the-shelf classifiers to give utterance classification performance that is surprisingly close to what can be achieved using conven- tional word-trigram recognition requiring man- ual transcription. In our method, unsupervised training is first used to train a phone n-gram model for a particular domain; the output of recognition with this model is then passed to a phone-string classifier. The classification ac- curacy of the method is evaluated on three dif- ferent spoken language system domains. 1 Introduction A major bottleneck in building data-driven speech pro- cessing applications is the need to manually transcribe training utterances into words. The resulting corpus of transcribed word strings is then used to train application- specific language models for speech recognition, and in some cases also to train the natural language components of the application. Some of these speech processing ap- plications make use of utterance classification, for exam- ple when assigning a call destination to naturally spoken user utterances (Gorin et al., 1997; Carpenter and Chu- Carroll, 1998), or as an initial step in converting speech to actions in spoken interfaces (Alshawi and Douglas, 2001). In this paper we present an approach to utterance clas- sification that avoids the manual effort of transcribing training utterances into word strings. Instead, only the desired utterance class needs to be associated with each sample utterance. The method combines automatic train- ing of application-specific phonotactic models together with token sequence classifiers. The accuracy of this phone-string utterance classification method turns out to be surprisingly close to what can be achieved by conven- tional methods involving word-trigram language mod- els that require manual transcription. To quantify this, we present empirical accuracy results from three differ- ent call-routing applications comparing our method with conventional utterance classification using word-trigram recognition. Previous work at AT&T on utterance classification without words used information theoretic metrics to dis- cover “acoustic morphemes” from untranscribed utter- ances paired with routing destinations (Gorin et al., 1999; Levit et al., 2001; Petrovska-Delacretaz et al., 2000). However, that approach has so far proved impractical: the major obstacle to practical utility was the low run- time detection rate of acoustic morphemes discovered during training. This led to a high false rejection rate (be- tween 40% and 50% for 1-best recognition output) when a word-based classification algorithm (the one described by Wright et. al (1997)) was applied to the detected se- quence of acoustic morphemes. More generally, previous work using phone string (or phone-lattice) recognition has concentrated on tasks in- volving retrieval of audio or video (Jones et al., 1996; Foote et al., 1997; Ng and Zue, 1998; Choi et al., 1999). In those tasks, performance of phone-based systems was not comparable to the accuracy obtainable from word- based systems, but rather the rationale was avoiding the …

Content truncated to 500 words