manassehzw/sna-dataset-annotated

An annotated, speaker-relabelled, and loudness-normalised Shona (sna) speech dataset prepared through a reproducible Modal-based data engineering pipeline.

This release addresses speaker label contamination in the original source labels by replacing identity columns with acoustically-derived speaker assignments.

Why this annotated release exists

The original source speaker labels are contaminated (multiple voices assigned to the same identity). This release replaces source identity labels with programmatically derived speaker clusters and rederived gender labels.

Pre/Post relabelling snapshot

• Pre-classification clips: 16980
• Pre-classification unique source speakers: 133
• Post-classification speaker clusters: 46
• Noise clips dropped after rescue: 1741
• Noise clips rescued by centroid similarity: 794

Processing Summary

1. Speaker relabelling using ECAPA embeddings + HDBSCAN + noise rescue.
2. Gender relabelling using a Shona-calibrated Logistic Regression classifier trained on ECAPA embeddings.
3. Noise drop (cluster_id == -1) and schema rebuild.
4. Loudness normalisation to -23 LUFS (EBU R128) with clipping protection.
5. Speaker-stratified split into train/validation/test.

Relabelling Method

• Speaker embeddings: speechbrain/spkrec-ecapa-voxceleb (192-d)
• Clustering: HDBSCAN (min_cluster_size=50, min_samples=10, metric euclidean, method eom)
• Noise rescue: cosine similarity threshold 0.75
• Gender model: Logistic Regression on L2-normalised ECAPA embeddings
  • Training clips (female): 160
  • Training clips (male): 152
  • Train accuracy: 1.0
  • CV accuracy: 1.0

Loudness Normalisation

• Target: -23 LUFS
• Skip tolerance: +/-1 LU
• Post-gain protection: hard clip to [-1.0, 1.0]
• Input LUFS mean/std: -22.659 / 5.301
• Output LUFS mean/std: -22.999 / 0.243

Splits

• Train: 12170
• Validation: 1504
• Test: 1565

Split strategy is speaker-stratified by clip proportion, preserving speaker distribution across splits.

Data Fields

• audio: 24kHz mono float audio (loudness-normalised)
• transcription: normalized Shona transcription
• source_id: original clip identifier from source dataset
• speaker_id: acoustically-derived speaker class id
• speaker_clip_count: clip count for the assigned speaker_id
• language: language code (sna)
• gender: cluster-level resolved label (Female / Male / Unknown)
• has_punctuation: punctuation indicator from normalized transcript
• snr_db: signal-to-noise proxy metric
• speech_ratio: fraction of VAD frames classified as speech
• quality_score: composite quality metric
• duration: clip duration in seconds
• speaker_assignment_confidence: confidence for speaker assignment

Uses

Direct use

• Shona ASR model training and adaptation
• TTS subset construction by filtering on speaker and quality metadata
• Speech data quality analysis and dataset curation workflows

Out-of-scope

• Identity verification / forensic use without additional validation
• Demographic representativeness claims without dedicated study

Bias, Risks, and Limitations

• Inherits source demographic/dialect distribution.
• Relabelled speaker IDs are acoustic clusters, not identity-verified persons.
• Confidence values are useful for filtering, not absolute truth scores.
• Some residual label uncertainty can remain in ambiguous/noisy clips.

Citation

If you use this dataset, cite both this release and the source dataset:

@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}
}