Fast and Accurate OOV Decoder on High-Level Features

Abstract

This work proposes a novel approach to out-of-vocabulary (OOV) keyword search (KWS) task. The proposed approach is based on using high-level features from an automatic speech recognition (ASR) system, so called phoneme posterior based (PPB) features, for decoding. These features are obtained by calculating time-dependent phoneme posterior probabilities from word lattices, followed by their smoothing. For the PPB features we developed a special novel very fast, simple and efficient OOV decoder. Experimental results are presented on the Georgian language from the IARPA Babel Program, which was the test language in the OpenKWS 2016 evaluation campaign. The results show that in terms of maximum term weighted value (MTWV) metric and computational speed, for single ASR systems, the proposed approach significantly outperforms the state-of-the-art approach based on using in-vocabulary proxies for OOV keywords in the indexed database. The comparison of the two OOV KWS approaches on the fusion results of the nine different ASR systems demonstrates that the proposed OOV decoder outperforms the proxy-based approach in terms of MTWV metric given the comparable processing speed. Other important advantages of the OOV decoder include extremely low memory consumption and simplicity of its implementation and parameter optimization.

Dynamic Acoustic Unit Augmentation with BPE-Dropout for Low-Resource End-to-End Speech Recognition

“…The aforementioned problems of OOV words handling and low-resource data conditions need to be addressed when building an ASR system. If the system is a conventional hybrid (HMM-DNN-based acoustic model and word-based n-gram language model), the OOV problem is often solved by dynamically expanding the system’s vocabulary and/or adapting the language model (e.g., [ 11 , 12 , 13 ]). A less common approach is to use a subword-based n-gram language model [ 14 ].…”

Multimodal Word Discovery and Retrieval with Phone Sequence and Image Concepts

“…The BBN system [10] combined several acoustic models based on DNN, LSTM and CNN on subword units to perform joint decoding and handled OOV queries on the sub-word unit. The STC keyword search system [11] combined 9 different acoustic models based on DNN and GMM with a phone-posterior based OOV decoder [12].…”

Dynamic Acoustic Unit Augmentation With BPE-Dropout for Low-Resource End-to-End Speech Recognition

“…If the system is a conventional hybrid (HMM-DNN-based acoustic model and word-based ngram language model), the OOV problem is often solved by dynamic expanding the system's vocabulary and/or adapting the language model (e.g. Khokhlov et al, 2017;Gandhe et al, 2018;Malkovsky et al, 2020). A less common approach is to use a subword-based n-gram language model (Smit et al, 2017).…”