Release Date: 11/20/2026
Format: JSONL, WAV
Size: 910.17 MB
The Saraiki ASR Dataset is a 9.5h resource designed to support commercial development in speech and language technologies for the Saraiki language. It contains utterance-level audios and corresponding transcriptions, extracted from ELAN files. This dataset aims to advance the development of commercial AI and NLP technologies for the underrepresented Saraiki language while supporting language documentation, preservation, and future multilingual research.
Pricing details
Your purchase is the license to the raw data. Once purchased, you're responsible for storing and using this dataset.
Your purchase includes a license to the data, paid directly to the dataset vendor, and a 5% ($100.00) platform fee paid to MDC.
Read More About The PoliciesLicensing
MDC Data Licence Agreement 1.0 (MDC-DLA-1.0)
https://community.mozilladatacollective.com/mdc-data-licence-agreement-1-0/You must not attempt to identify, re-identify or expose the identity of any speaker in this dataset. This includes cross-referencing with other datasets, voice-print matching, or any other technique intended to link recordings to real people.
You must not attempt to identify, re-identify or expose the identity of any speaker in this dataset. This includes cross-referencing with other datasets, voice-print matching, or any other technique intended to link recordings to real people.
Segmented field recordings of spoken Saraiki, for automatic speech recognition: Saraiki audio → Saraiki text.
| Task | automatic speech recognition: Saraiki audio → Saraiki text |
| Speech language | Saraiki (skr) |
| Segments | 5836 |
| Audio | 9.45 hours, 16000 Hz mono 16-bit WAV |
| Recordings | 129, recorded 2026-02-19 to 2026-02-28 (10 recording days) |
| Licence | MDC Data Licence Agreement 1.0 (see LICENCE.md) |
| split | segments | hours | recordings |
|---|---|---|---|
| train | 4573 | 7.45 | 101 |
| dev | 665 | 1.02 | 15 |
| test | 598 | 0.98 | 13 |
| total | 5836 | 9.45 | 129 |
Splits are made by recording: every segment of a recording is in the same split, so the test set contains no audio from recordings seen in training. Target proportions were 0.8/0.1/0.1 of audio duration (seed 13).
saraiki-asr/
├── README.md
├── LICENCE.md
├── stats.json
├── train.jsonl dev.jsonl test.jsonl
└── clips/<recording>/<id>.wav
Each line of a .jsonl file describes one clip:
| field | description |
|---|---|
id | unique segment id: <session>_<tier>_<index> |
audio_filepath | path of the clip, relative to the dataset folder |
duration | clip length in seconds |
text | transcription in Saraiki, Arabic script, Unicode NFC |
src_lang | ISO 639-3 code of the speech |
speaker | speaker label scoped to one recording, e.g. 20260223-skr001:A |
session | annotation file the segment comes from (recordings can have parts -1, -2) |
recording | recording id; splits are made on this field |
The recordings were annotated in ELAN. Each annotation on the transcription tier (tx_skr_native, or Tier-0 in files where that tier is empty) became one clip, cut from the original recording at its annotated start and end and resampled to 16000 Hz mono.
Text processing: Unicode NFC normalisation, whitespace collapsed, bracketed notes such as [laugh] removed. Annotations with no letters (placeholders such as .......) were dropped.
Clips shorter than 0.5 s or longer than 30.0 s were dropped.
from datasets import load_dataset, Audio
ds = load_dataset("json", data_dir="saraiki-asr")
ds = ds.map(lambda x: {"audio": "saraiki-asr/" + x["audio_filepath"]})
ds = ds.cast_column("audio", Audio(sampling_rate=16000))
The .jsonl files are also valid NeMo manifests (audio_filepath, duration, text).
Licensed under the MDC Data Licence Agreement 1.0. See LICENCE.md and https://community.mozilladatacollective.com/mdc-data-licence-agreement-1-0/.
This dataset belongs to the MDC Datasets series of speech resources for underrepresented languages. It contains utterance-level Saraiki audio with corresponding text (Saraiki transcriptions, and English translations in the speech translation dataset), extracted from ELAN annotation files. Its aim is to advance AI and NLP technologies for the underrepresented Saraiki language while supporting language documentation, preservation and future multilingual research.
Steward: MDC Datasets, Mozilla Data Collective
You must not attempt to identify, re-identify or expose the identity of any speaker in this dataset. This includes cross-referencing with other datasets, voice-print matching, or any other technique intended to link recordings to real people.
You might also like...
date | recording date, YYYYMMDD |
start / end | position of the clip in the original recording, seconds |
overlap | true if another annotated speaker talks during the clip |
tier | annotation tier the transcription came from |
source_media | file name of the original recording |
Dataset attributes