An integrated language identification for code- switched speech using decoded-phonemes and support vector machine

Abstract

Automatic language identification (LID) is a specialized area of Human Language Technology in which the language(s) used in spoken utterances are identified and correctly classified given a predetermined number of targeted languages. Currently, most multilingual speakers have the ability and tendency for engaging in code-switching -a mixed-language phenomenon that is referred to as the usage of more than one language in utterances. This paper presents the proposed scheme for automatic language identification integrated with an automatic speech recognition system to identify languages used in a mixed-language speech context. The front-end speech recognition system feeds the decoded phonemes into the LID system. We used hidden Markov models to build acoustic models of a combined phoneme set that handles multiple languages within an utterance. A spoken utterance is converted into feature vectors with attributes that represents the statistical occurrences of each acoustic units. A supervised support vector machine (SVM) technique is trained with feature vector sequences of phoneme units. The back-end SVM classifier based on n-gramstructures is used to classify/identify the phoneme feature vectors. We conducted experiments with two commonly mixed Northern Sotho and English telephone-based speech corpora. The experimental results showed that, by using shared phonemic vowels in the combined phoneme set, the word error rate (WER) was reduced with 3.6%. Moreover, the proposed approach yields significantly acceptable performance with language identification rate of 85.0% on code-switched speech corpus.

Investigating Sentiment-Bearing Words- and Emoji-based Distant Supervision Approaches for Sentiment Analysis

“…The Nguni group is the largest group which includes Seswati, isiNdebele, isiXhosa, and isiZulu. In this instance, our study focuses on the second-largest group-the Sotho-Tswana group comprising Sepedi (Northern Sotho) (Mabokela and Manamela, 2013), Sesotho (Southern Sotho), and Setswana (Statista, 2022).…”

OLR 2021 Challenge: Datasets, Rules and Baselines

“…One kind of multilingual speech recognition a These authors contributed equally to this work. system adopts the cascade architecture, which first determines the language through an LID system and then transcribe the speech using the monolingual ASR system [5], [6]. The other kind of multilingual speech recognition system uses an end-to-end holistic architecture that collects data from all languages and trains a single model that can handle languages identification and speech recognition altogether [7].…”