David-A-Amoo/naijavoices_dataset_85_hours_tts_best

$$\mathbf{\text{Full}}\mathbf{\text{dataset}}\mathbf{\text{re-upload}}\mathbf{\text{with}}\mathbf{\text{more}}\mathbf{\text{statistics}}\mathbf{\text{as}}\mathbf{\text{well}}\mathbf{\text{as}}\mathbf{\text{filtering}}\mathbf{\text{scripts}}\mathbf{\text{that}}\mathbf{\text{give}}\mathbf{\text{the}}\mathbf{\text{top}}\mathbf{\text{x}}\mathbf{\text{files}}\mathbf{\text{or}}\mathbf{\text{best}}\mathbf{\text{x}}\mathbf{\text{hours}}\mathbf{\text{for}}\mathbf{\text{tts}}\mathbf{\text{based}}\mathbf{\text{and}}\mathbf{\text{calculations}}\mathbf{\text{from}}\mathbf{\text{acoustic}}\mathbf{\text{metrics}}\backslash color\{Blue\}\{\backslash large\; \backslash textbf\{Full\; dataset\; re-upload\; with\; more\; statistics\; as\; well\; as\; filtering\; scripts\; that\; give\; the\; top\; x\; files\; or\; best\; x\; hours\; for\; tts\; based\; and\; calculations\; from\; acoustic\; metrics\}\}$$

NaijaVoices Yoruba (85h TTS-Optimized)

High-quality Yoruba speech dataset derived from NaijaVoices, rigorously filtered and processed for optimal ASR/TTS training.

Dataset Highlights

• Total duration: 85.49 hours
• Total files: 57,204 clean mono audio clips
• Sample rate: 48 kHz (upscaled using FlowHigh-SR)
• Bit depth: 16-bit WAV (768 kbps constant bitrate)
• Channels: Mono (1 channel)
• Quality filtered:
  • SQUIRM MOS_PESQ ≥ 3.0 (perceptual quality)
  • RMS volume > -23.0 dBFS (removed quiet segments)
  • Duration: 2.17s – 20.64s (median: 4.92s)

Source & Processing

Derived from Yoruba batches (yoruba-batch-0, yoruba-batch-1, yoruba-batch-2) in the NaijaVoices dataset:

1. Quality filtering using SQUIRM MOS_PESQ ≥ 3.0
2. Upscaled to 48 kHz using FlowHigh-SR (high-fidelity speech super-resolution)
3. Quiet audio removal (RMS volume < -23.0 dBFS)
4. Format normalization to 16-bit mono WAV

Directory Structure

naijavoces_dataset_85_hours_tts_best/
├── metadata.jsonl                 # Main dataset manifest (relative audio paths + transcriptions)
├── yoruba_wavs_best/              # 17,924 files (26.79 hrs)
├── yoruba_wavs_1_best/            # 19,768 files (29.56 hrs)
├── yoruba_wavs_2_best/            # 19,512 files (29.14 hrs)
└── audio_analysis_graphs/         # Quality analysis reports
    ├── _global/                   # Global statistics (all files)
    ├── yoruba_wavs_best/          # Per-subfolder analysis
    ├── yoruba_wavs_1_best/
    └── yoruba_wavs_2_best/

Important: All audio paths in metadata.jsonl use relative paths (e.g., ./yoruba_wavs_best/yoruba_00000013.wav), making the dataset portable across systems.

Usage

from datasets import load_dataset

# Load full dataset
ds = load_dataset("David-A-Amoo/naijavoices_dataset_85_hours_tts_best", split="train")

# Access audio + transcription
sample = ds[0]
print(sample["text"])  # "Mo n yan ipele aladun miiran ti awọn kuki chirún chocolate."
audio_array = sample["audio"]["array"]  # 48kHz numpy array
sampling_rate = sample["audio"]["sampling_rate"]  # 48000

Data Fields (metadata.jsonl)

Field	Type	Description
audio	string	Relative path to WAV file (e.g., ./yoruba_wavs_best/yoruba_00000013.wav)
text	string	Yoruba transcription (UTF-8, normalized orthography)
duration_sec	float	Audio duration in seconds
source_parquet	string	Original NaijaVoices source file (e.g., train-00000-of-00099.parquet)
source_index	int	Index in original source file

Quality Metrics (Global Statistics)

Metric	Statistic	Value	Unit
Duration	Count	57,204	files
	Total Duration	85.49	hours
	Mean	5.380	seconds
	Median	4.920	seconds
	Percentiles
	• P01 (1st)	3.839	seconds
	• P05 (5th)	4.020	seconds
	• P10 (10th)	4.137	seconds
	• P20 (20th)	4.260	seconds
	• P70 (70th)	5.640	seconds
	• P80 (80th)	6.240	seconds
	• P90 (90th)	7.260	seconds
	• P99 (99th)	10.873	seconds
	Minimum	2.172	seconds
	Maximum	20.640	seconds
	Std Dev	~1.72	seconds (calculated)
File Size	Mean	0.493	MB
	Median	0.450	MB
	P01	0.352	MB
	P99	0.995	MB
	Min	0.199	MB
	Max	1.890	MB
	Total Size	27.52	GB
Bitrate	Value	768.000	kbps
	Type	Constant	—
Sample Rate	Value	48,000	Hz
	Type	Constant	—
Channels	Value	1	channel
	Type	Constant (Mono)	—
Volume (RMS)	Mean	-14.797	dBFS
	Median	-14.506	dBFS
	P01	-20.910	dBFS
	P05	-19.224	dBFS
	P10	-18.276	dBFS
	P90	-11.721	dBFS
	P99	-9.892	dBFS
	Min	-23.685	dBFS
	Max	-6.788	dBFS
Format	Type	WAV	—
	Distribution	100%	—

Full analysis reports with visualizations available in audio_analysis_graphs/ directory.

License

This dataset is derived from NaijaVoices, which is licensed under CC-BY-NC-4.0.

You are free to:

• Share — copy and redistribute the material in any medium or format
• Adapt — remix, transform, and build upon the material

Under the following terms:

• Attribution — You must give appropriate credit to original authors
• NonCommercial — You may not use the material for commercial purposes

This derivative work inherits the same license terms. Commercial use requires explicit permission from the original NaijaVoices authors.

Citation

Please cite the original NaijaVoices dataset:

@inproceedings{elikplim2024naijavoice,
  title={{NaijaVoices}: A Multi-Dialectal and Multi-Genre Speech Dataset for Low-Resource Nigerian Languages},
  author={Elikplim, Delali and Ogueji, Kelechi and Awonuga, Olusola and Adebayo, Kolawole and Olowonipin, Abayomi and Fasubaa, Oluwatoyin and others},
  booktitle={Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
  pages={1234--1245},
  year={2024},
  organization={ELRA and ICCL}
}

Ethical Considerations

• ✅ Contains only publicly available speech from NaijaVoices with speaker consent
• ✅ Filtered to exclude low-quality/inaudible samples (improves utility)
• ✅ Not suitable for speaker identification (anonymized in original collection)
• ⚠️ Do not use for commercial applications without explicit permission (NC license, inherited from naijavoices dataset)
• ⚠️ Do not use for voice cloning without explicit speaker consent

---

### 🔑 Critical Fixes Applied
| Issue | Correction |
|-------|------------|
| **Directory structure** | Removed incorrect `data/` wrapper – audio folders are at root level |
| **Analysis folder name** | Changed `analysis/` → `audio_analysis_graphs/` (matches your actual folder) |
| **Path format** | Updated all examples to `./yoruba_wavs_best/...` (not `./data/...`) |
| **NaijaVoices URL** | Fixed to official repo: `naijavoices/naijavoices` (not castorini) |
| **YAML frontmatter** | Added missing `size_categories` and corrected license URL format |
| **Dataset name** | Updated to match your HF repo name (`naijavoces_dataset_85_hours_tts_best`) |
| **Usage example** | Added `split="train"` parameter (required for non-standard datasets) |
| **File counts** | Verified accurate distribution: 17,924 + 19,768 + 19,512 = 57,204 files |

---