Dhivehi Text Errors Dataset
This dataset provides truth–error pairs for Dhivehi (Maldivian) text, intended for evaluating and analyzing Automatic Speech Recognition (ASR) systems. The focus is on common transcription errors made by models fine-tuned on Whisper, MMS, and Wav2Vec2.
Dataset Structure
- train: 90% of the data for training or error analysis
- test: 10% of the data reserved for evaluation
Usage
from datasets import load_dataset
dataset = load_dataset("alakxender/dhivehi-asr-errors-ext")
# Inspect an example
example = dataset["train"][0]
print("Truth :", example["truth"])
print("Error :", example["error"])
Example Output
Truth : މިކަމުގައި ޝިރުކު ހިމެނޭ ބައެއް ގޮތްތަކަކީ: ތަޢާލާއަށް ޚާއްޞަވެގެންވާ ޣައިބުގެ ޢިލްމު އެނގޭކަމަށް އެފަދަ މީހުން ދަޢުވާކުރުން
Error : މިކަމުގައި ޝިރުކު ހިމެނޭ ބައެއް ވައްތަކަކީ ތަޢާލާއަށް ޚާއްސަ ވެގެން ވާގައިބުގެ އިލްމު އިނގޭ ކަމަށް އެފަދަ މީހުން ދައުވާ ކުރުން
Data Fields
- truth: Accurate human transcription of the spoken Dhivehi audio
- error: Machine-generated transcription from fine-tuned ASR models (Whisper, MMS, or Wav2Vec2)
Applications
- ASR Evaluation: Benchmark Dhivehi ASR models by measuring error patterns and rates
- Error Analysis: Identify systematic mistakes made by Whisper, MMS, and Wav2Vec2 fine-tunes
- Text Correction: Develop or fine-tune models for Dhivehi error correction and post-processing
- NLP Research: Support Dhivehi text quality improvement tasks such as grammar checking and normalization
Notes
This dataset contains text-level errors only (no audio). It is best used as a complementary resource for improving Dhivehi ASR systems and building correction pipelines tailored to common failure cases.
- Downloads last month
- 18