Comparative Study of Tokenization Algorithms for End-to-End Open Vocabulary Keyword Detection

The advent of Deep-Learning techniques and the increasing importance of personalization in voice assistants fueled the need for open vocabulary keyword detection systems, in which, the user can enroll a keyword using audio or text as a modality. A text enrollment-based custom-keyword detection system has to detect that the input speech signal is matched with enrolled keyword-phrase or not. The methods used for tokenization of keyword phrase can alter the custom-keyword detection performance. Hence, in this study, we explore and evaluate different tokenization methods which includes, phoneme-level tokenization, character-level tokenization, and subword tokenization (Byte Pair Encoding, and Unigram). The efficiency of these methods is studied using an end-to-end custom-keyword detection framework, on five different datasets. The findings of this study reveal valuable insights into the suitability of different tokenization methods for custom-keyword detection task. From results, it is observed that proposed architecture with phoneme-based tokenization shows best detection accuracy of 99.54%, 98.12% and 90.05% on Libriphrase, Qualcomm, Google speech command datasets, respectively.