Nepali Repetitions & Formatted Numbers Dataset (nep_reps_)

This dataset is specifically designed for fine-tuning Nepali Text-to-Speech (TTS) models, such as Chatterbox, to prevent attention alignment drift, loop repetitions, and tail hallucinations when processing phonetically similar words, tongue twisters, and consecutive digits.

Dataset Structure

The dataset contains a single .parquet split with the following columns:

• audio: A dictionary containing:
  • bytes: Raw WAV audio data (mono, 24kHz).
  • path: The local filename of the audio clip.
• text: The corresponding normalized text in Devanagari script (expanded numbers like "एक, दुई, तीन" instead of raw digits).
• language: The language identifier ("ne" for Nepali).
• duration: The length of the audio clip in seconds.

Data Subsets

1. Repeated Words & Tongue Twisters (126 samples): Focuses on phonetically similar word pairs and traditional tongue twisters.
2. Numbers & Formats (646 samples): Focuses on sequential counting, telephone numbers, bank account sequences, currencies, and dates.

Usage

You can load this dataset directly using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("Firoj112/nep_reps_")
print(dataset["train"][0])