dsfsi-anv/multilingual-nchlt-dataset

NCHLT Auxiliary Speech Corpus - Combined Multilingual Dataset

Dataset Description

This is a combined multilingual version of the NCHLT Auxiliary Speech Corpus, compiled by the Data Science for Social Impact (DSFSI) research group at the University of Pretoria to facilitate easier benchmarking and multi-language speech recognition research.

The original auxiliary data was collected during the National Centre for Human Language Technology (NCHLT) project for the 11 official languages of South Africa, and was made publicly available by SADiLaR in 2019. This combined dataset brings together all 11 language datasets into a unified format.

Dataset Summary

This dataset provides between 20 and 170 hours of speech data per language along with orthographic transcriptions, totaling over 1,420 hours of speech across all languages. The data was originally collected using a smartphone application called Woefzela and includes recordings from approximately 3,400 speakers across all languages.

Original Source: Individual language datasets are available from SADiLaR (South African Centre for Digital Language Resources).

DSFSI Contribution: Combined the separate language datasets into a unified multilingual corpus for easier benchmarking and cross-lingual research.

Supported Languages

The dataset covers all 11 official languages of South Africa:

• Afrikaans (afr)
• South African English (eng)
• isiNdebele (nbl)
• isiXhosa (xho)
• isiZulu (zul)
• Sepedi (nso)
• Sesotho (sot)
• Setswana (tsn)
• Siswati (ssw)
• Tshivenda (ven)
• Xitsonga (tso)

Dataset Structure

Data Instances

Each recording consists of:

• Audio file in WAVE format (16-bit, mono, PCM, 16kHz)
• Orthographic transcription
• Speaker metadata (age, gender, location)
• Language identifier
• Confidence scores (PDP scores)
• Duration information

Data Fields

Corpus Level:

• language: ISO 639-3 language code

Speaker Level:

• speaker_id: Anonymous numeric speaker identifier
• age: Numeric age value (majority between 18-55)
• gender: "male" or "female"
• location: South African province where recorded

Recording Level:

• audio: Audio file path/data
• md5sum: MD5 checksum for consistency verification
• duration: Recording duration in seconds
• pdp_score: Phone-based dynamic programming confidence score
• transcription: Orthographic transcription (prompt text)

Data Splits

The auxiliary data consists of two subsets per language:

Aux1: Recordings from NCHLT-baseline not included in the official 2014 release

• Total duration: 780.6 hours
• Total speakers: 2,220
• Total utterances: 689,343

Aux2: Recordings from raw data pool after NCHLT-baseline selection

• Total duration: 640.7 hours
• Total speakers: 1,181
• Total utterances: 583,551

Dataset Statistics by Language

Language	Aux1 Utterances	Aux1 Hours	Aux1 Speakers	Aux2 Utterances	Aux2 Hours	Aux2 Speakers	Total Hours
Afrikaans (afr)	51,666	42.7	210	46,934	39.1	94	81.8
English (eng)	42,006	29.8	210	54,091	38.9	113	68.7
isiNdebele (nbl)	34,445	42.6	148	96,200	120.1	208	162.7
Sepedi (nso)	62,965	64.9	210	52,371	51.8	105	116.7
Sesotho (sot)	68,599	73.9	210	47,238	43.5	98	117.4
Siswati (ssw)	60,238	78.4	197	126,932	167.0	226	245.4
Setswana (tsn)	67,702	70.2	210	34,800	37.0	75	107.2
Xitsonga (tso)	67,530	83.7	198	827	0.7	6	84.4
Tshivenda (ven)	78,009	93.7	208	43,446	54.9	86	148.6
isiXhosa (xho)	81,821	103.0	209	50,720	55.0	107	158.0
isiZulu (zul)	74,362	97.9	210	29,992	32.7	63	130.6
Total	689,343	780.6	2,220	583,551	640.7	1,181	1,421.3

Dataset Creation

Curation by DSFSI

This combined dataset was curated by the Data Science for Social Impact (DSFSI) research group at the University of Pretoria to:

• Facilitate multilingual and cross-lingual ASR research
• Enable easier benchmarking across South African languages
• Provide a unified format for training multilingual models
• Support research in under-resourced language technologies

Original Data Collection

Data was collected through a dedicated campaign using the Woefzela smartphone application, developed specifically for the NCHLT project.

Data Collection Process

• Participants read utterances displayed on mobile devices
• Recording environments were less controlled than studio settings
• Quality checks implemented on-device in semi-realtime:
  • Clipping detection
  • Volume detection
  • Speech cutting detection
  • Speech sufficiency metrics
• Balance maintained between:
  • Male and female speakers (approximately 50:50)
  • Rural and urban pronunciations
  • Age range (majority 18-55 years)

Quality Measures

• Language proficiency verified by qualified language practitioners
• Phone-based dynamic programming (PDP) confidence scores used to rank recordings
• Automatic rescheduling of recordings that failed quality measures

Annotations

Orthographic transcriptions were created for all utterances as part of the data collection process.

Licensing Information

This dataset is released under the CC BY (Creative Commons Attribution) license, consistent with the original SADiLaR release.

Citation Information

If using this combined DSFSI dataset, please cite:

1. This combined dataset (please add appropriate DSFSI citation when available)

2. The original auxiliary corpus paper:

@article{badenhorst2022nchlt,
  title={NCHLT Auxiliary speech data for ASR technology development in South Africa},
  author={Badenhorst, Jaco and de Wet, Febe},
  journal={Data in Brief},
  volume={41},
  pages={107860},
  year={2022},
  publisher={Elsevier},
  doi={10.1016/j.dib.2022.107860}
}

3. Related research on the data:

@article{badenhorst2019usefulness,
  title={The usefulness of imperfect speech data for ASR development in low-resource languages},
  author={Badenhorst, Jaco and de Wet, Febe},
  journal={Information},
  volume={10},
  number={9},
  pages={268},
  year={2019},
  doi={10.3390/info10090268}
}

4. The original NCHLT Speech Corpus:

@inproceedings{barnard2014nchlt,
  title={The NCHLT speech corpus of the South African languages},
  author={Barnard, Etienne and Davel, Marelie H and van Heerden, Charl and de Wet, Febe and Badenhorst, Jaco},
  booktitle={Proceedings of the Workshop on Spoken Language Technologies for Under-Resourced Languages},
  year={2014},
  address={St Petersburg, Russia}
}

Original Data Sources

SADiLaR Repositories

The original individual language datasets are available from SADiLaR (South African Centre for Digital Language Resources):

• Afrikaans: https://hdl.handle.net/20.500.12185/522
• English: https://hdl.handle.net/20.500.12185/523
• isiNdebele: https://hdl.handle.net/20.500.12185/513
• Sepedi: https://hdl.handle.net/20.500.12185/518
• Sesotho: https://hdl.handle.net/20.500.12185/519
• Siswati: https://hdl.handle.net/20.500.12185/515
• Setswana: https://hdl.handle.net/20.500.12185/520
• Xitsonga: https://hdl.handle.net/20.500.12185/521
• Tshivenda: https://hdl.handle.net/20.500.12185/516
• isiXhosa: https://hdl.handle.net/20.500.12185/514
• isiZulu: https://hdl.handle.net/20.500.12185/517

Additional Information

Value of the Data

• Scale: Provides over 1,420 hours of speech data across 11 languages
• Diversity: Includes varied acoustic conditions and speaker demographics
• Under-resourced languages: Critical resource for South African languages with limited technology support
• Proven utility: Shown to improve ASR system recognition accuracy when combined with existing data
• Multilingual research: Enables cross-lingual and multilingual model development

Use Cases

The NCHLT resources have been successfully applied in:

• AwezaMed: Speech-to-speech translation mobile application for maternal healthcare communication
• Broadcast harvesting: Automatic transcription and data collection from broadcast media
• ASR development: Training and benchmarking automatic speech recognition systems
• Multilingual model training: Cross-lingual transfer learning and language comparison studies

Known Limitations

• Speaker demographics may be biased toward university campuses (collection sites)
• Recording quality varies due to mobile collection environment
• Some languages (e.g., Xitsonga Aux2) have significantly less data than others
• Some speaker overlap exists between Aux1, Aux2, and the original NCHLT test sets (documented in original paper)

Ethics Statement

• All participants provided informed consent for recording and corpus inclusion
• Personal identifying information (names, national IDs, phone numbers) removed prior to publication
• Only anonymized speaker identifiers released with data
• Data collection and release protocols approved for research use

Acknowledgments

Original Data Collection:

• Supported by South African Department of Arts and Culture (NCHLT project)
• Compilation enabled by Department of Science and Innovation through SADiLaR

Combined Dataset:

• Compiled by Data Science for Social Impact (DSFSI) research group
• University of Pretoria

Contact

For questions about:

• This combined dataset: Contact DSFSI research group
• Original data: Refer to SADiLaR or the original publication

Contributing

If you use this dataset and identify issues or have suggestions for improvement, please open an issue on the dataset repository.

---

Maintained by: Data Science for Social Impact (DSFSI), University of Pretoria
Original Data Provider: SADiLaR (South African Centre for Digital Language Resources)
Original Collectors: CSIR and Stellenbosch University (NCHLT Project)