AI-Based Regional Bangla Speech to Standard Bangla and English Translation System

Description

This dataset contains 5,200 WAV audio files with Bangla regional speech transcripts. The data is prepared to support machine learning and natural language processing research, especially Bangla automatic speech recognition, regional dialect identification, transcript normalization, and translation-based language analysis. The dataset represents spoken Bangla from eight district-based regional categories: Barishal, Chittagong, Habiganj, Narail, Narsingdi, Rangpur, Sylhet, and Tangail. Each district category contains 650 audio files. The dataset is balanced across all eight districts, resulting in a total of 5,200 audio samples. The original dataset is divided into training, validation, and testing sets. For each district, 520 files are used for training, 65 files are used for validation, and 65 files are used for testing. This creates a total split of 4,160 training files, 520 validation files, and 520 testing files. The file naming convention follows a district-wise serial format: district_serial.wav. For example, Barishal files are named from barishal_0001.wav to barishal_0650.wav. The same naming pattern is used for all district categories. The dataset metadata is maintained in spreadsheet form with file names and corresponding regional transcripts. Additional columns such as standard Bangla and English translation may also be included for normalization and translation tasks. The transcripts reflect natural spoken Bangla and may contain regional vocabulary, pronunciation-based spellings, conversational markers, incomplete sentences, fillers, and code-mixed words. These characteristics make the dataset useful for studying real-world speech and dialect variation in Bangla. Since the data is based on speech, careful transcript verification and audio-transcript alignment are recommended before model training.

Institutions

• Southeast University

  Dhaka Division, Dhaka

Categories

Licence