Shona Speech Dataset (sna-dataset)

A cleaned, metadata-rich Shona (sna) speech dataset prepared through a reproducible data engineering pipeline for downstream ASR and TTS workflows.

This release is intended as a general-purpose standard corpus: quality metadata is provided, but aggressive opinionated filtering is avoided so users can apply task-specific thresholds.

Dataset Details

Dataset Sources

• Source dataset: google/WaxalNLP
• Subset used: sna_asr (train + validation + test labeled splits)
• Pipeline repository: sna-data-pipeline (Modal-based end-to-end processing)

Annotation Disclaimer

Some metadata fields (especially source_speaker_id and gender) are inherited from the original source dataset may contain labeling errors. This release preserves source-provided identity annotations for provenance and reproducibility. While quality-control and audit checks were applied, metadata assignments are not guaranteed to be 100% correct. Future revisions may include improved speaker-consistency and sex/gender label audits, with any corrections documented in the changelog.

Uses

Direct Use

• Automatic Speech Recognition (ASR) training and fine-tuning
• Speech-text data analysis in Shona
• TTS subset construction via metadata filters (e.g., speaker and quality constraints)

Out-of-Scope Use

• Clinical, forensic, or speaker verification use without additional validation
• Claims of broad dialect/sociolinguistic representativeness without dedicated bias study

Dataset Structure

• train: 13,532 clips
• validation: 1,640 clips
• test: 1,808 clips

The split is speaker-stratified by clip proportion using speaker_idx (not speaker-disjoint), so each split retains similar speaker distribution while preserving clip diversity.

Data Fields

• audio: 24kHz mono float audio
• transcription: normalized Shona transcription
• source_id: original clip identifier from source dataset
• source_speaker_id: original speaker hash from source dataset
• speaker_idx: stable integer speaker index (frequency-sorted at ingest)
• speaker_clip_count: number of clips for that speaker in the cleaned dataset
• language: normalized language code (sna)
• gender: normalized gender label (Male / Female)
• has_punctuation: whether transcript contains sentence punctuation
• snr_db: signal-to-noise proxy metric
• speech_ratio: fraction of VAD frames classified as speech
• quality_score: composite metric (snr_db with reliability penalties)
• duration: post-processing clip duration in seconds

Dataset Creation

Curation Rationale

The goal was to produce a robust, well-documented Shona speech corpus that preserves provenance, adds high-value metadata, and remains broadly usable for both ASR and TTS downstream tasks.

Data Processing Pipeline

Source Code

The dataset was produced with a reproducible, Modal-based data engineering pipeline implemented in Python (ingest -> metadata annotation -> text normalization -> audio normalization -> cleanup -> split/upload).
Pipeline source code: sna-data-pipeline

1. Ingest

   • Loaded labeled sna_asr splits from WaxalNLP
   • Preserved provenance by renaming:
     • id → source_id
     • speaker_id → source_speaker_id
   • Assigned stable speaker_idx

2. Metadata Annotation

   • Normalized language and gender
   • Added speaker_clip_count

3. Text Normalization

   • Standardized punctuation/quotes/dashes
   • Cleaned unsupported symbols
   • Added has_punctuation

4. Audio Normalization

   • Resampled to 24kHz mono
   • WebRTC VAD trimming with smoothing
   • Internal silence-gap compression
   • Computed snr_db, speech_ratio, quality_score, duration
   • Dropped zero-speech/empty clips and blacklisted speaker clips

5. Post-normalization Cleanup

   • Dropped very short clips (duration threshold)
   • Dropped singleton-speaker rows
   • Refreshed speaker_clip_count

6. Split and Upload

   • Speaker-stratified 80/10/10 split
   • Published as sna-dataset with dataset card + audit trail

Bias, Risks, and Limitations

• Inherits demographic/dialectal distribution from source data.
• Not speaker-disjoint across splits (same speaker can appear in multiple splits).
• Quality metrics are useful for ranking/filtering, but not absolute perceptual guarantees.
• Users should apply task-specific filtering for strict TTS quality targets.

Citation

Prepared by Manasseh Changachirere (Harare Institute of Technology).

If you use this dataset, please cite the original source:

@inproceedings{niang2024waxalnlp,
  title={WaxalNLP: A Large Scale High Quality Speech Dataset for African Languages},
  author={Niang, El Hadj Mamadou and Dieng, Moustapha and Ba, Thierno Ibrahima and Ndiaye, Mamadou Boumedine and others},
  booktitle={Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
  year={2024}
}