A Synchronized Chittagonian Speech Corpus for Dialect ASR and Speech-to-Speech Translation

The corpus contains 8,004 Chittagonian speech samples paired with corresponding Chittagonian transcripts and Standard Bangla translations. The textual component of the dataset is taken from the publicly available Kothon dataset, a Chittagonian–Standard Bangla parallel text resource.

In this work, the original text pairs were extended with newly collected speech recordings from native Chittagonian speakers, creating a synchronized speech-text corpus for automatic speech recognition (ASR), speech-to-speech translation, machine translation, dialect normalization, and text-to-speech (TTS) . The dataset is organized into predefined training, validation, and test splits containing 6,403, 801, and 800 samples, respectively.