Subspace-Based Representation and Learning for Phonotactic Spoken Language Recognition

Abstract

Phonotactic constraints can be employed to distinguish languages by representing a speech utterance as a multinomial distribution or phone events. In the present study, we propose a new learning mechanism based on subspacebased representation, which can extract concealed phonotactic structures from utterances, for language verification and dialect/accent identification. The framework mainly involves two successive parts. The first part involves subspace construction. Specifically, it decodes each utterance into a sequence of vectors filled with phone-posteriors and transforms the vector sequence into a linear orthogonal subspace based on low-rank matrix factorization or dynamic linear modeling. The second part involves subspace learning based on kernel machines, such as support vector machines and the newly developed subspace-based neural networks (SNNs). The input layer of SNNs is specifically designed for the sample represented by subspaces. The topology ensures that the same output can be derived from identical subspaces by modifying the conventional feed-forward pass to fit the mathematical definition of subspace similarity. Evaluated on the "General LR" test of NIST LRE 2007, the proposed method achieved up to 52%, 46%, 56%, and 27% relative reductions in equal error rates over the sequence-based PPR-LM, PPR-VSM, and PPR-IVEC methods and the lattice-based PPR-LM method, respectively. Furthermore, on the dialect/accent identification task of NIST LRE 2009, the SNN-based system performed better than the aforementioned four baseline methods.

Spoken Language Recognization Based on Features and Classification Methods

“…Hung-Shin Lee et al, [2] the suggested technique obtained relative error rate reductions of 52 percent, 46 percent, 56 percent, and 27 percent when compared to the sequence-based PPR-LM, PPR-VSM, and PPR-IVEC methods, as well as the lattice-based PPR-LM method. They offer a novel learning technique for language verification and dialect/accent recognition that is based on subspace-based representation and can extract hidden phonotactic features from utterances.…”

Detecting English Speech in the Air Traffic Control Voice Communication

“…We decided to explore the English language detection (ELD) based on an ASR output for the following two reasons: 1) there is about 20% of non-English speech, so running ASR on all the data does not bring a large overhead (compared to filtering non-English prior to the ASR block), 2) standard acoustic-based Language IDentification (LID) can be too data-consuming [2,3,4] and its accuracy suffers in bi-lingual code-switching environments [5]. The ASR-based ELD systems share similarities with phonotactic LID sytems [6,7,8,9]: the former relies on word confusion networks, the latter relies on phoneme sequences.…”

“…We used a vocabulary of 28.4k unique tokens. Out of this, 15.3k tokens are 5-letter way-points from [38], and 5.2k tokens are airline designators for call-signs 6,7 . Pronunciations were generated using Phonetisaurus tool [39] with a G2P model trained on Librispeech lexicon [40].…”

A comparison of cepstral and spectral features using recurrent neural network for spoken language identification

“…Artificial intelligence has achieved new goals in developing intelligent algorithms for processing languages (both text and audio), making humans more interactive with the systems. Many languages have been computationally modelled with a lot of effort [1][2][3]. In the early days of voice recognition and speaker recognition, speech signals were portrayed as a normal (smooth input without first-and second-order derivatives) input to systems.…”