Mahidol University Releases Free Thai Speech-to-Text Models for Commercial Use

·3 views

Researchers from the Biomedical and Data Lab at Mahidol University have released Thonburian Whisper, a suite of Thai-language automatic speech recognition models fine-tuned from OpenAI's Whisper. The models are available in multiple sizes, ranging from small variants that run on standard hardware to a large-v3 version requiring a GPU, achieving Word Error Rates of 6.59 and 7.42 respectively on the Common Voice Thai test set. All models are licensed under Apache 2.0, making them freely usable for commercial applications without requiring additional permission. Distilled versions offering faster inference are also available for developers prioritising speed over maximum accuracy. The project was originally published in 2022 and the models are hosted publicly on Hugging Face.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Amazon Athena Lets Users Query S3 Data with SQL, No Servers Required

Amazon Athena is a serverless interactive query service from AWS that enables users to analyze data stored in Amazon S3 using standard SQL, without setting up or managing any infrastructure. Users simply upload their dataset to an S3 bucket, define the data schema in Athena, and run SQL queries that are processed in parallel across the cloud. The service supports multiple file formats including CSV, JSON, ORC, Avro, and Apache Parquet, making it flexible for various data types. Athena is particularly suited for ad-hoc analysis, log exploration, and querying data lakes, eliminating the need for dedicated servers or database clusters. Its serverless model and familiar SQL interface make it accessible for students and small teams who need scalable analytics without complex infrastructure overhead.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Dual-Orchestration Pattern Proposed to Separate Machine and Human Workflows in Banking

A technical paper published on DEV Community proposes a dual-orchestration architecture for modern core banking platforms facing conflicting execution demands. The approach, called the Xenon Architecture Standards, pairs Temporal for automated, low-latency distributed transaction workflows with Flowable for multi-week human approval and review processes. The design is mapped to Banking Industry Architecture Network (BIAN) service domains, assigning orchestration responsibility based on state duration, throughput, and human involvement. The paper argues that using a single orchestration engine for both profiles causes problems such as database bloat, thread starvation, and compromised audit compliance. Code examples illustrate how Temporal handles compensating transaction Sagas, including fund reservation, sanctions screening, and account crediting.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AWS CodeDeploy Automates App Deployments Across EC2, Lambda, and ECS

AWS CodeDeploy is a fully managed deployment service from Amazon Web Services that automates the transfer of application code, scripts, and configuration files to various computing environments. It supports Amazon EC2 servers, AWS Lambda functions, and Amazon ECS containers, retrieving application revisions from sources like Amazon S3 and GitHub. The service eliminates manual deployment steps by using a configuration file called appspec.yml to define lifecycle hooks that guide each deployment consistently. CodeDeploy offers both in-place and blue/green deployment strategies, giving developers flexibility to balance speed against risk. It also integrates with CloudWatch to monitor deployment health and can automatically roll back to a previous stable version if a failure is detected.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

JaiTTS: Free Thai Voice-Cloning Model Beats Human Benchmark on Short-Text Accuracy

Jasmine Technology Solution (JTS) released JaiTTS-v1.0, a Thai text-to-speech voice-cloning model, publishing their findings on arXiv on April 30, 2026, alongside open-access code and a demo. The model achieved a Character Error Rate (CER) of 1.94% on short-text tasks, narrowly surpassing the human reference score of 1.98% — though it fell short of human performance on longer texts. Built on a modified VoxCPM autoregressive architecture and trained on a large Thai-focused audio corpus, JaiTTS can handle mixed Thai-English text and numerals without requiring text normalization. In a blind listening test involving 20 native Thai evaluators, JaiTTS outperformed two commercial systems across 283 of 400 compared pairs, while losing in 58 pairs. The model also processes audio roughly nine times faster than real time, and is available for free use.

0 comments Read more at DEV Community