Exploring Questions, Creating Answers

Current state-of-the-art (SOTA) Automatic Speech Recognition (ASR) models are multilingual. While these models have greatly improved transcription accuracy for high-resourced languages, under-resourced languages still require further language specific optimizations and finetuning to achieve acceptable levels of accuracy. In this work we explore ways of improving ASR for Dhivehi, an under-resourced South Asian language, by finetuning pretrained multilingual ASR models, Subword Modelling, Language Model (LM) decoding and Automatic Spelling Correction. We finetune 5 Dhivehi ASR models and apply our accuracy boosting techniques, with one of our models achieving a new state-of-the-art Word Error Rate (WER) of 14.26% on the Dhivehi Common Voice ASR benchmark, which is a 31.93% relative WER improvement over the existing SOTA of 20.95%. We create a new Dhivehi text corpus, and train 2 new Dhivehi LMs to support our accuracy boosting techniques.

This article summarizes the results of a 2023 salary survey of Software Developers in the Maldives, conducted to assess industry-wide compensation levels. Based on 125 responses (≈12.9% of the estimated developer population), the average monthly salary was MVR 24,734.70, with a median of MVR 20,000 and an estimated ±8% margin of error. The analysis shows clear salary differences across sectors, with the private sector offering the highest average pay, followed by state-owned enterprises and government roles. Salary increases strongly with years of experience, while education shows a non-linear relationship, suggesting that either strong practical skills or higher academic qualifications lead to better earnings. These findings provide a useful benchmark for employers and job seekers in the Maldivian software industry.