Data Engine (DE) System
The DE System is an NLU engine (and associated developer toolkit) for providing robust, low-cost, natural-speech call-routing for greater than 20-25 routing destinations via a data-driven SLM (statistical language model) methodology. Compared to conventional approaches, it lowers development and maintenance costs of natural-speech, say-anything call-routing grammars while increasing their routing accuracy and call completion rates.
DE Mechanism
The DE System follows a four-step process: (a) preparation of training data consisting of transcriptions and their corresponding routing destination labels; (b) automatic extraction of relevant keywords (and/or key-phrases) from training data for recognizer and NLU grammar lexicons; (c) automatic generation of natural-speech, say-anything recognizer and NLU grammars through combination of extracted keywords with a unique domain and speaker-independent garbage model; and (d) testing and tuning of recognizer and NLU grammars. The recognizer grammar is loaded inside the third-party recognizer and allows detection of specified keywords in a wide variety of natural-speech, say-anything utterances. The NLU grammar is loaded inside Vestec NLU engine and processes the output of the third-party recognizer for semantic interpretation.
DE Benefits
The DE System offers powerful advantages over conventional SLM grammar approaches for natural-speech, say-anything call-routing applications. It reduces development and maintenance costs via: (1) smaller training data-set requirement on account of focus on relevant keywords; (2) faster grammar development due to fully transparent and manipulable lexicons; and (3) rapid grammar tuning through automated error analysis and corrective suggestions. In addition, the DE System increases routing accuracy and call-completion rates via: (4) better processing of out-of-grammar utterances due to robust garbage model; (5) superior contextual interpretation of utterances on account of sophisticated nBest processing; and (6) better handling of accented speakers through recovery of “lost” keywords.
