A CNN-MFCC baseline for isolated-word recognition in Bishnupriya Manipuri: corpus development and speaker-independent system evaluation

Life Cycle Reliability and Safety Engineering · Springer Nature · 2026-10-14 · DOI: 10.1007/s10772-026-10302-8

Аннотация

Bishnupriya Manipuri, an Indo-Aryan language, is spoken by more than 120,000 people in Northeast India and Bangladesh and has been designated as a vulnerable language by UNESCO. This language lacks an annotated voice corpus, and no automatic speech recognition system applying deep learning has been established for it. This article confronts both shortcomings by two main contributions: (1) Established a read speech corpus consisting of 2,800 utterances, featuring 28 words and 10 sentences, recorded from 10 native BM speakers within the 20 to 35 age range; and (2) a Convolutional Neural Network-based isolated-word recognizer, trained on 20-dimensional Mel Frequency Cepstral Coefficient attributes, achieving a word recognition accuracy of 92.8% with a 70:30 data split ratio. Because this convention allows a model to exploit speaker-specific cues rather than word-general acoustic structure, we additionally evaluate the system under a genuine speaker-independent protocol: 10-fold leave-one-speaker-out (LOSO) cross-validation, in which every fold trains on nine speakers and tests exclusively on the tenth, previously unseen speaker. Under this protocol, WAcc falls to 75.2% $$\pm$$ 8.6% (mean $$\pm$$ standard deviation across the ten held-out speakers). Gaussian Mixture Model (GMM) and Support Vector Machine (SVM) baselines trained on identical features degrade far more severely under the same speaker-independent protocol (39.3% $$\pm$$ 9.5% and 37.1% $$\pm$$ 7.4% respectively), indicating that classical models rely substantially on speaker-specific characteristics, whereas the CNN’s convolutional features generalise measurably better to unseen speakers. We report per-class precision, recall and F1, a confusion analysis, and training/validation loss curves obtained under the speaker-independent protocol, and discuss the corpus’s limitations - scale, age range, dialect coverage, and closed-vocabulary scope. In order to facilitate further studies on this language with limited resources, all corpus materials will also be rendered publicly available.

Читать у издателя →

Похожие статьи