CoolFace
Datasetpublic

suleiman2003/W_hausa_v3

Cleaned Hausa Speech Dataset v3 A cleaned and processed Hausa speech dataset built from multiple open-source Hugging Face datasets. Dataset Description This dataset contains cleaned, normalized, and deduplicated Hausa speech audio with aligned transcriptions. All audio is: Sample rate: 16,000 Hz (mono) Format: FLAC (lossless, embedded in Parquet) Duration range: 1–30 seconds per clip Loudness normalized: -20 dBFS RMS VAD trimmed: Non-speech segments removed with… See the full description on the dataset page: https://huggingface.co/datasets/suleiman2003/W_hausa_v3.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face
0likes327downloads
Dataset Card

Cleaned Hausa Speech Dataset v3

A cleaned and processed Hausa speech dataset built from multiple open-source Hugging Face datasets.

Dataset Description

This dataset contains cleaned, normalized, and deduplicated Hausa speech audio with aligned transcriptions. All audio is:

  • —Sample rate: 16,000 Hz (mono)
  • —Format: FLAC (lossless, embedded in Parquet)
  • —Duration range: 1–30 seconds per clip
  • —Loudness normalized: -20 dBFS RMS
  • —VAD trimmed: Non-speech segments removed with crossfaded joins
  • —Silence trimmed: Leading/trailing silence removed (25 dB threshold)
  • —Deduplicated: By audio content hash and (text, speaker_id) pairs

Sources

SourceDescription
google/WaxalNLPHausa TTS data from WaxalNLP
naijavoices/naijavoices-datasetNaijaVoices Hausa batches (0, 1, 2)
vpetukhov/bible_tts_hausaBible TTS Hausa (single speaker)

Dataset Structure

Features

  • —audio: Audio waveform (FLAC, 16kHz mono)
  • —text: Hausa transcription (NFC-normalized Unicode)
  • —speaker_id: Integer speaker identifier (unique per source × speaker)

Usage

python
from datasets import load_dataset

dataset = load_dataset("suleiman2003/W_hausa_v3", split="train")

# Play first sample
print(dataset[0]["text"])
print(dataset[0]["audio"]["sampling_rate"])  # 16000

Processing Pipeline

  1. 1.Source Parquet files downloaded from HuggingFace Hub
  2. 2.Audio decoded, resampled to 16kHz mono
  3. 3.Silence trimming (25 dB threshold)
  4. 4.WebRTC VAD trimming (aggressiveness=2) with 10ms crossfade
  5. 5.RMS loudness normalization to -20 dBFS
  6. 6.Quality filters: duration (1-30s), clipping ratio (<1%)
  7. 7.Deduplication by audio hash and (text, speaker_id)
  8. 8.FLAC encoding and Parquet sharding