Work

The First Character-Level Visual Speech Recognition for Bangla: data and code

Year: 2026

Type: software

Abstract: Pre-print: https://doi.org/10.21203/rs.3.rs-11207165/v1Prediction files, training record, and scoring code behind the first character-level visual speech recognition work for Bangla. Five training seeds of a visual-only character transducer on 14,431 LipBengal test clips from 30 held-out speakers: 78.15 +/- 1.45% charWER, 63.48 +/- 0.79% character error over Unicode code points and 71.73 +/- 0.98% over grapheme clusters, against a 99.79% majority-class floor. The included script recomputes all three rates from the prediction files.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.23036344

Authors: Uddin, Abdul Hasib, Arif, Abu Shamim Mohammad

Institution: Khulna University

Language: English

Topic: Speech and Audio Processing

Subfield: Signal Processing

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7214746006

Cites:

Cited by:

Related to: 10