The Odyssey SA Voice Corpus (V0.4) is a 1000-hour multilingual South African speech dataset spanning 8 languages, designed for automatic speech recognition (ASR), code-switching research, and large audio model (LAM) evaluation.

This repository provides a 69-minute evaluation preview (Episode 001 — Nobantu Vilakazi). The preview mirrors the segmentation, metadata schema, validation pipeline, and structural standards used in the full commercial release.

This subset is intended for benchmarking and research evaluation purposes.

Dataset Composition (Evaluation Preview)

Episode ID: ep_001_nobantu_vilakazi Total Clips: 138 Total Duration: 4,140.0 seconds (69.0 minutes)

Clip Classification

Premium: 138

Standard: 0

Discarded: 0

All clips in this preview meet V0.4 quality thresholds.

ASR Confidence Metrics

Mean ASR Confidence (duration-weighted): 0.91

Median ASR Confidence: 0.92

Confidence scores were generated using an automated ASR validation pipeline and verified during dataset generation.

Audio & Linguistic Characteristics

Language Distribution (duration-weighted):

en-ZA: 82.5%

zu-ZA: 17.5%

Code-Switching Rate: 38.4% of clips Short-Form Clips: 12 Clips Requiring Human Verification: 0.0%

Audio consists of natural, spontaneous South African conversational speech recorded in controlled studio conditions. Each clip represents an atomic speech unit with a verbatim transcript and structured metadata.

Quality & Validation

Blocking validation issues: 0

Manifest integrity checks: Passed

Schema validation: Passed

Deterministic indexing enforced

This preview reflects the same structured pipeline used in the full 1000-hour release.

Data Collection & Consent

All participants provided documented consent for research and commercial dataset distribution. Personal identifying information has been excluded. Audio was recorded in controlled environments and processed through a structured segmentation and validation pipeline.

Intended Use

This evaluation preview is intended for:

ASR benchmarking

Code-switching research

Accent adaptation experiments

Multilingual audio model testing

This subset is not intended for large-scale model training.

Access to the full Odyssey SA Voice Corpus (V0.4) may be available for research or commercial licensing upon request.

Directory Structure odyssey_voice_V0.4_sample/ ├─ README.md ├─ license.txt ├─ manifest.csv ├─ consent_inventory.csv ├─ audio/ ├─ per_clip_cards/ └─ transcripts_full/

Loading Example from datasets import load_dataset

dataset = load_dataset("ODYSSEYAILABS/odyssey-sa-voice-v0.4-sample") print(dataset)

Citation

Odyssey AI Labs. Odyssey SA Voice Corpus (V0.4), Evaluation Preview.

Downloads last month: 16

Total file size:

129 MB