kawshikbuet17/bengali-telecom-customer-care-speech

Bengali Telecom Customer Care Synthetic Speech Dataset

Dataset Description

This dataset contains synthetic Bengali speech generated from telecom and customer-care style text prompts.

The dataset is intended for experiments with:

• Bengali ASR/STT
• Bengali TTS
• Speech-to-text preprocessing
• Telecom/customer-care domain adaptation
• Synthetic speech research

Important Disclosure

This is a synthetic speech dataset generated using the OmniVoice TTS system in voice-cloning mode. A real female voice recording and its transcript were used as the reference voice for cloning.

Generation settings: bfloat16 precision, num_step=16 diffusion sampling steps, and speed=1.0 speaking-rate control.

All audio in this dataset is synthetic. It does not contain real customer-care recordings, real customer conversations, or real user audio.

Dataset Size

Total examples: 10000

Approximate total duration: 26.82 hours

Sample rate: 24000 Hz

Splits:

Split	Examples
Train	9000
Validation	500
Test	500

Dataset Structure

bengali-telecom-customer-care-speech/
└── data/
    ├── train/
    │   ├── metadata.jsonl
    │   └── audios/
    ├── validation/
    │   ├── metadata.jsonl
    │   └── audios/
    └── test/
        ├── metadata.jsonl
        └── audios/

Each metadata row contains:

{
  "file_name": "audios/example.wav",
  "text": "Original Bengali text.",
  "text_normalized": "Normalized Bengali text.",
  "language": "bn",
  "is_synthetic": true,
  "sample_rate": 24000,
  "duration_sec": 5.32
}

Field Descriptions

Field	Description
file_name	Relative path to the audio file
text	Original Bengali text used for TTS generation
text_normalized	Normalized Bengali text suitable for ASR/STT training
language	Language code, bn for Bengali
is_synthetic	Whether the sample is synthetic
sample_rate	Audio sample rate
duration_sec	Audio duration in seconds

Intended Use

For TTS training, use:

text -> audio

For ASR/STT training, use:

audio -> text_normalized

How to Load / Download

Install the Hugging Face datasets library with audio support:

pip install -U "datasets[audio]"

Then load the dataset in Python:

from datasets import load_dataset

dataset = load_dataset("kawshikbuet17/bengali-telecom-customer-care-speech")

# Example: print first training sample
print(dataset["train"][0])

Limitations

Because this dataset is synthetic, models trained only on this dataset may not fully generalize to real-world speech, noisy phone calls, spontaneous conversation, different microphones, real accents, or background noise.

For production-quality ASR or TTS, this dataset should ideally be combined with real human-recorded Bengali speech.

License

License: cc-by-4.0

Prepared By

• Kawshik Kumar Paul
  Dept of CSE, BUET
  kawshikbuet17@gmail.com
• Nafiul Alam Fuji
  Dept of CSE, BUET
  nafiul.fuji@gmail.com