Datasets:

IAmNotAnanth
/

sinhala-ctc-111h

pretty_name: Sinhala ASR – OpenSLR SLR52 Consolidated (111h)
license: cc-by-4.0
language:
  - si
task_categories:
  - automatic-speech-recognition
tasks:
  - speech-recognition
tags:
  - sinhala
  - indic-asr
  - speech
  - ctc
  - nemo
  - silero-vad
  - openslr
  - low-resource
size_categories:
  - 100h-1k

Sinhala ASR – Consolidated OpenSLR (SLR52)

Dataset Summary

This dataset is a consolidated and cleaned version of the Sinhala Automatic Speech Recognition (ASR) dataset from OpenSLR (SLR52).

The original OpenSLR release distributes the data across multiple subsets (0–9, a–f). This repository merges all subsets into a single unified dataset containing approximately 111 hours of speech audio.

Dataset Description

Consolidation

All OpenSLR SLR52 subsets (0–9, a–f) have been merged into one dataset.
The resulting dataset contains ~111 hours of audio data.

Language Analysis

The dataset was analyzed and found to consist exclusively of:
- Sinhala-only utterances
- English-only utterances
No code-mixed utterances were observed.
All English-only samples were removed.

Duration Filtering

Audio samples shorter than 0.5 seconds or longer than 30 seconds were removed.

Silence Removal

Silero VAD was used to remove leading and trailing silence from all audio files.

After preprocessing, the dataset was reduced from approximately 185k original utterances to ~178k cleaned samples.

Audio Characteristics

Format: WAV
Sampling Rate: 16 kHz
Channels: Mono
Silence: Leading and trailing silence removed using VAD

No additional SNR-based filtering was required.

Text Processing

Transcriptions were found to be clean and well-formed
No text normalization beyond basic Unicode normalization was required

Dataset Structure

.
├── wavs/
│   ├── si_0000001.wav
│   ├── si_0000002.wav
│   └── ...
├── manifest.jsonl
└── README.md

Manifest Format (`manifest.jsonl`)

Each line is a JSON object with the following fields:

{
  "audio_filepath": "wavs/si_0000001.wav",
  "duration": 2.34,
  "text": "සිංහල වාක්‍යයක්",
  "lang": "si"
}

This format is compatible with ASR toolkits such as:

NVIDIA NeMo
ESPnet
Fairseq
Hugging Face Datasets

Intended Use

Automatic Speech Recognition (ASR)
CTC-based acoustic model training
Low-resource and Indic language research

Source Data

Original Dataset: OpenSLR – SLR52 (Sinhala ASR)
URL: https://openslr.org/52/

License

Please refer to the original OpenSLR SLR52 dataset license. Users should ensure compliance with the original data usage terms.

Citation

If you use this dataset, please cite the original OpenSLR SLR52 release.