babs/yoruba_lagos_nw

Yoruba Lagos–NWU (Internal)

Dataset Summary

Internal XTTS‑ready version of the Lagos–NWU Yoruba speech corpus (University of Lagos & North‑West University). The dataset contains single‑utterance WAV clips with Yoruba orthographic transcriptions including tone marks. Transcriptions are normalized to NFC while preserving tone information (combining marks are retained if present).

Important: This dataset is internal and not for public distribution. If uploading to Hugging Face, use a private repository and ensure you have permission to share.

Supported Tasks

• Text‑to‑speech (TTS)

Languages

• Yoruba (yo)

Data Source

• Local source path: /Users/emeke/Downloads/corpus
• Original transcriptions: utts.data.orig (tone‑marked, Unicode NFD)

Dataset Statistics

• Speakers: 33 (16 female, 17 male)
• Utterances: 4288 (train 4090, eval 198)
• Total audio duration: ~2.75 hours
• Skipped: 28 (too short)
• Tone diacritics preserved (NFC); combining marks may remain

Data Format

yoruba_lagos_nw/
  wavs/
    *.wav
  metadata_train_yo.csv
  metadata_eval_yo.csv
  metadata_train_yo.pipe
  metadata_eval_yo.pipe
  metadata.csv
  metadata.pipe
  summary.json
  README.md

Metadata

• metadata_train_yo.csv / metadata_eval_yo.csv (HF‑friendly CSV with header):

  audio,text,speaker_id
  

• metadata.csv (HF‑friendly CSV with header):

  audio,text,speaker_id,gender,split
  

• metadata_train_yo.pipe / metadata_eval_yo.pipe (Coqui‑style, pipe‑delimited for XTTS training):

  wavs/<file>.wav|<text>|<speaker_id>
  

• metadata.pipe (pipe‑delimited master metadata with header):

  audio|text|speaker_id|gender|split
  

Speaker IDs

<gender>_<speaker> (e.g., female_001, male_027)

Audio

• Format: WAV, PCM16, mono
• Sample rate: 22,050 Hz (resampled from 16 kHz)
• RMS normalized

Text Normalization

• Unicode NFC
• Control characters removed
• | replaced to avoid delimiter collisions
• Whitespace collapsed
• Tone marks preserved (critical for Yoruba pronunciation)

Quality Filters Applied

• Duration: 1.0–15.0 s
• Text rate: 1–30 chars/sec
• Silence ratio and clipping thresholds
• RMS floor to exclude overly quiet audio

Known Issues

• Some recordings have room reverb (per original corpus notes). Consider denoising if required.

License / Access

Internal dataset. Do not distribute without approval. Use private storage on Hugging Face.