Researchers release synthetic Bengali speech dataset for telecom customer-care scenarios

Researchers from arXiv:2608.20346 introduced a synthetic Bengali speech dataset tailored for telecom customer-care scenarios. The dataset comprises 10,000 audio-text pairs totaling approximately 26.82 hours of speech at a 24 kHz sampling rate. It includes predefined train, validation, and test splits of 9,000, 500, and 500 examples, respectively.

The speech was generated using OmniVoice in voice-cloning mode, leveraging a real female reference recording and transcript. The generation process used bfloat16 precision, 16 diffusion sampling steps, and a speaking-rate control value of 1.0. In addition to the original Bengali text, the dataset provides a normalized transcript field designed to support ASR/STT training and evaluation.

To assess text-audio consistency, the authors conducted an automatic intelligibility check over all 10,000 samples using a domain-adapted Whisper ASR model fine-tuned from bengaliAI/tugstugi_bengaliai-regional-asr_whisper-medium. They also performed a manual listening check on selected samples. The evaluation reported an average word error rate (WER) of 2.54% and an average character error rate (CER) of 0.59%, with median WER and CER values of 0.00%.

The dataset is publicly released on Hugging Face under the CC-BY-4.0 license, enabling broader access for research and development in Bengali speech technologies.