Building and Benchmarking a Cascaded STT–TTS Pipeline for Jeseri-to-Malayalam Speech Translation Using Pretrained Transformer Models
Computer and Decision Making: An International Journal · 14.11.2026 · 4 (1) · 328-344
DOI: 10.1922/comdem.v4i1.19135 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computer and Decision Making: An International Journal, online erschienen am 14.11.2026. Verfasst von Bijimol T. K. und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Malayalam is the language of Kerala and it has regional forms or dialects. At present, 13 regional forms or dialects are in operation, and Jeseri is one among them spoken in the Lakshadweep Islands, influenced by Arabic due to the historical advent of Islam. Jeseri is a zero resource language, making it difficult to preserve and teach. The proposed system is a speech-to-speech translation system for Jeseri-to-Malayalam created with the help of the manually annotated parallel data. An inventory of 3,000 simple Malayalam-Jeseri pairs of sentences was generated using 6,000 occurrences of commonly used Malayalam sentences, ansd eight native Jeseri speakers in Kadmat Island were recorded. Among them, 2,700 samples were used for training and 300 for testing. The system includes Jeseri speech input; Speech-to-Text (STT) conversion based on refined SeamlessM4T (Massively Multilingual & Multimodal Machine Translation) and Whisper-large models; Text-to-Speech (TTS) conversion based on Google Text-to-Speech (gTTS), Indic-Parler-TTS, and OpenAI TTS-1. Word Error Rate (WER), Character Error Rate (CER), and Mean Opinion Score (MOS) were used in assessing performance. Findings indicate that for STT conversion, the optimized SeamlessM4T model with the minimum WER of 0.3800 and CER of 0.1840 outperforms the Whisper-large model. For the case of TTS conversion, Google TTS outperforms both Indic-Parler-TTS and OpenAI TTS-1 with lowest WER (0.060) and CER (0.050) and the highest MOS (4.80). Overall, the cascaded STT-TTS pipeline is an efficient tool in supporting Jeseri-to-Malayalam STS translation and has high performance based on the significant decreases in WER and CER.
Zitieren
APA
K., B. T., & Lee, I. (2026). Building and Benchmarking a Cascaded STT–TTS Pipeline for Jeseri-to-Malayalam Speech Translation Using Pretrained Transformer Models. Computer and Decision Making: An International Journal, 4(1), 328-344. https://doi.org/10.1922/comdem.v4i1.19135
Vancouver
K. BT, Lee I. Building and Benchmarking a Cascaded STT–TTS Pipeline for Jeseri-to-Malayalam Speech Translation Using Pretrained Transformer Models. Computer and Decision Making: An International Journal. 2026;4(1):328-344. doi:10.1922/comdem.v4i1.19135
BibTeX
@article{K2026,
author = {K., Bijimol T. and Lee, Ickjai},
title = {Building and Benchmarking a Cascaded STT–TTS Pipeline for Jeseri-to-Malayalam Speech Translation Using Pretrained Transformer Models},
journal = {Computer and Decision Making: An International Journal},
year = {2026},
volume = {4},
number = {1},
pages = {328--344},
doi = {10.1922/comdem.v4i1.19135}
}
EUVIMED