CoolFace
Datasetpublic

iamTangsang/nepali_to_english_pipeline_evaluation

Nepali-English Speech-to-Text Translation Evaluation Dataset Dataset Description This dataset is designed for evaluating Nepali→English speech-to-text translation pipelines. It contains audio recordings of 300 Nepali sentences, spoken by three speakers, covering a range of sentence types (statements, questions, commands, complex sentences, and named entities/numbers). Each sentence is paired with: Source text (Nepali) transcription Reference English translation… See the full description on the dataset page: https://huggingface.co/datasets/iamTangsang/nepali_to_english_pipeline_evaluation.

sourceHugging Facemitupdated 1y agoView on Hugging Face
0likes12downloads
Dataset Card

Nepali-English Speech-to-Text Translation Evaluation Dataset

Dataset Description

This dataset is designed for evaluating Nepali→English speech-to-text translation pipelines. It contains audio recordings of 300 Nepali sentences, spoken by three speakers, covering a range of sentence types (statements, questions, commands, complex sentences, and named entities/numbers).

Each sentence is paired with:

  • —Source text (Nepali) transcription
  • —Reference English translation
  • —Audio recording at 16 kHz, mono .wav format
  • —Metadata including speaker_id and category

Motivation

The dataset is intended for system evaluation only (not model training). It can be used to assess:

  • —Automatic Speech Recognition (ASR) accuracy (WER, CER)
  • —Machine Translation (BLEU, COMET, etc.)
  • —Punctuation restoration performance
  • —Speaker robustness in end-to-end pipelines

Dataset Structure

  • —Speakers: 3 (speaker01, speaker02, speaker_03)
  • —Sentences per speaker: 300
  • —Total audio clips: 900
  • —Sampling rate: 16 kHz mono
  • —Languages:
  • —Nepali (source text)
  • —English (reference translation)

Categories

  • —Statements (50%)
  • —Questions (20%)
  • —Commands (10%)
  • —Complex sentences (10%)
  • —Named entities / numbers (10%)

Example Data

python
from datasets import load_dataset, Audio

ds = load_dataset("iamTangsang/pipeline_evaluation", split="evaluation")
ds = ds.cast_column("audio_path", Audio(sampling_rate=16_000))

print(ds[0])
python
{
  'id': 'speaker_01_statements_1',
  'speaker_id': 'speaker_01',
  'category': 'statements',
  'audio_path': AudioDecoder Object,
  'source_text': 'यो किताब हो।',
  'reference_translation': 'This is a book.'
}

Usage

Global evaluation

python
# Full dataset
all_data = ds

Per-speaker evaluation

python
s1 = ds.filter(lambda x: x["speaker_id"] == "speaker_01")
s2 = ds.filter(lambda x: x["speaker_id"] == "speaker_02")
s3 = ds.filter(lambda x: x["speaker_id"] == "speaker_03")

Intended Uses

  • —Evaluate ASR models for Nepali
  • —Benchmark translation systems Nepali → English
  • —Study speaker variability in speech-to-text translation
  • —Provide testbed for end-to-end speech translation pipelines

Not intended for training models — only for evaluation.

Citation

If you use this dataset, please cite:

ruby@dataset{your_username_2025_nepali_english_eval,
  title     = {Nepali-English Speech Translation Evaluation Dataset},
  author    = {iamTangsang, praneshp1org},
  year      = {2025},
  publisher = {Hugging Face Datasets},
  url       = {https://huggingface.co/datasets/your-username/pipeline_evaluation}
}