Improving Indian Spoken-Language Identification by Feature Selection in Duration Mismatch Framework

Abstract

Paper presents novel duration normalized feature selection technique and two-step modified hierarchical classifier to improve the accuracy of spoken language identification (SLID) using Indian languages for duration mismatched condition. Feature selection averages random forest-based importance vectors of open SMILE features of different duration utterances. Although it improves the SLID system's accuracy for mismatched training and testing durations, the performance is significantly reduced for short-duration utterances. A cascade of inter-family and intra-family classifiers with an additional class to improve false language family estimation. All India Radio data set with nine Indian languages and different utterance durations was used as speech material. Experimental results showed that 150 optimal features with the proposed modified hierarchical classifier showed the highest accuracy of 96.9% and 84.4% for 30 s and 0.2 s utterances for the same train-test duration. However, we achieved an accuracy of 98.3% and 61.9% for 15 and 0.2 s test duration when trained with 30 s duration utterance. Comparative analysis showed a significant improvement in accuracy than several SLID systems in the literature.

An Overview of Indian Spoken Language Recognition from Machine Learning Perspective

“…Aarti et al in [2] provided an organized review about the diferent linguistic cues according to the major language families in India. Further, in the literature, hierarchical LID systems with a front-end language family recognizer are proposed for improved LID performance [11,76,160]. Bengali, Hindi, Punjabi, Tamil, Urdu…”

“…This corpus contains speech data of some rarely studied Indian languages, such as Angami, Bodo, Khasi, Hrangkhawl, and Sumi. Some of the other sources for availing standard Indian speech data are Speehocean, 11 Indic-TTS, 12 13 There are also developments in open-source corpora, such as Mozilla Common Voice, 14 OpenSLR, 15 with speech data for the Indian languages. In Table 3, we have summarized the key information about the major speech corpora developed for Indian spoken language recognition research.…”

“…Their proposed algorithm reduced the feature dimensions to only 150 optimized components that improved the LID accuracy among diferent duration mismatch test conditions. The same authors extended the approach in [11] with the cascading of inter-language-family and intra-language-family based LID models. The irst stage trained a binary classiier for recognizing the Indo-Aryan and Dravidian languages.…”

“…The front-end language family classiier can recognize these languages at the beginning, followed by a dedicated LID system. In the literature, there are few research attempts that adapted this approach [11,22,76,159,160]. Further exploration in this approach, especially with state-of-the-art neural architectures is promising.…”