Language – Agnostic, Limited Vocabulary Speech Recognition System
Non-Confidential description of the invention
A speech recognition system that recognizes words spoken by users in any language or in any combination of languages with high accuracy upon being presented with between 2-8 previous examples of the word having been spoken by the same users in the same ambient environment. The speech recognition system transforms input sound files into a sparse, distributed representation using neural networks intended to mimic the representational capacity of the human pre-frontal cortex. The sparsity of the representation in the transformed space permits accurate recognition of individual word tokens using machine learning methods, independent of the language used, the speaker’s accent or prevailing noise conditions. Speech recognition systems are commonly used to set reminders, make purchases, and dictate notes and messages. This is because state-of-the-art speech recognition systems rely on training on large, annotated samples of speech from users to attain acceptable performance (measured in word error rate, or WER). Such large corpora of speech exist predominantly in English, and there too for simple conversational speech. Our speech recognition suffers from no such restrictions, and thus can be used in a variety of settings not considered practical for existing speech recognition systems. Some such settings include retail point-of-sale, supply chain inventory and auditing, medical transcription and stock-checks, logistical planning and dispatching, etc.
Mr. Surajit Ghosh (M. Tech Student, CSE), Dr. Nisheeth Srivastava (CSE)
Date of Filing