Improving out of vocabulary words recognition accuracy for an end-to-end Russian speech recognition system

Abstract

Automatic Speech Recognition (ASR) systems are experiencing an active introduction into our daily lives, simplifying the way we interact with electronic devices. The advent of end-to-end approaches has only accelerated this process. However, the constant evolution and a high degree of inflection of the Russian language lead to the problem of recognizing new words outside the vocabulary (Out Of Vocabulary, OOV) because they did not take part in the training process of the ASR system. In such a case, the ASR model tends to predict the most similar word from the training data which leads to a recognition error. This is especially true for ASR models that use decoding based on a Weighted Finite State Transducer (WFST), since they are obviously limited by the list of vocabulary words that may appear as a result of recognition. In this paper, this problem is investigated on the basis of an open data set of the Russian language (common voice) and an integrated ASR system using the WFST decoder. A method for retraining an integral ASR system based on the discriminative loss function MMI (maximum mutual information) and a method for decoding the integral model using a TG graph are proposed. Discriminative learning allows smoothing the probability distribution of acoustic class prediction, thus adding more variability in the recognition results. Decoding using the TG graph, in turn, is not limited to recognizing only vocabulary words and allows the use of a language model trained on a large amount of external text data. An eight-hour subset from the common voice base is used as a test set. The total number of OOV words in this test sample is 18.1 %. The results show that the use of the proposed methods allows to reduce the word recognition error (Word Error Rate, WER) by 3 % in absolute value relative to the standard method of decoding integral models (beam search), while maintaining the ability to recognize OOV words at a comparable level. The use of the proposed methods should improve the overall quality of recognition of ASR systems and make such systems more resistant to the recognition of new words that were not involved in the learning process.

Keywords

• automatic speech recognition
• end-to-end asr
• discriminative training
• oov words
• weighted finite state transducer