README.md

metadata

language:
  - af
pretty_name: Afrikaans Speech Dataset for Whisper Fine-Tuning
tags:
  - automatic-speech-recognition
  - speech
  - audio
  - afrikaans
  - low-resource
  - multilingual
license: cc-by-4.0
task_categories:
  - automatic-speech-recognition
dataset_info:
  features:
    - name: audio_id
      dtype: string
    - name: chunk_index
      dtype: int32
    - name: transcript_word_count
      dtype: int32
    - name: transcript_char_count
      dtype: int32
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: transcript
      dtype: string
  splits:
    - name: train
      num_bytes: 5381293102
      num_examples: 5603
    - name: validation
      num_bytes: 578182293
      num_examples: 602
    - name: test
      num_bytes: 586814137
      num_examples: 611
  download_size: 6362237408
  dataset_size: 6546289532
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*

Afrikaans Speech Dataset for Whisper Fine-Tuning

Dataset Card

Dataset Summary

This dataset consists of approximately 56 hours of Afrikaans speech extracted from church sermons, paired with cleaned and aligned transcripts. It is specifically prepared for fine-tuning multilingual ASR models like OpenAI's Whisper (particularly large-v3) on low-resource Afrikaans speech

The audio is segmented into fixed 30-second chunks (with 3-second overlaps for context preservation) at 16 kHz mono 16-bit PCM. Transcripts are normalized using Whisper's non-English text standardization rules (lowercase, no punctuation/diacritics, bracket/parenthesis removal, whitespace collapse).

This dataset aims to improve Whisper's performance on real-world, spontaneous Afrikaans speech (e.g., sermons, announcements, conversations) with varied accents and noise.

• Language: Afrikaans (af)
• Domain: Informal/spontaneous speech, primarily South African religious/community content
• Total Hours: ~50 hours (post-processing)
• Chunks: Thousands of 30-second segments across train/validation/test splits (80/10/10)
• License: CC-BY-4.0 (intended for research purposes)

Supported Tasks and Leaderboards

• Task: Automatic Speech Recognition (ASR) fine-tuning
• Models: Optimized for Whisper (tiny to large-v3); compatible with any seq2seq ASR model
• Evaluation Metric: Word Error Rate (WER) – expect reductions vs. base Whisper on Afrikaans test sets

Languages

• Primary: Afrikaans (South African variant)
• Speakers may use English words in sentences (minimal)

Dataset Structure

Data Splits

Split	Hours (approx)	Chunks (approx)	Description
train	46.6	~5603	Training data
validation	5.0	~622	Validation (early stopping)
test	5.0	~606	Held-out evaluation

Data Instances

Each instance is a 30-second audio chunk with transcription

• audio_id: Original audio ID
• chunk_index: Sequential chunk number relative to raw audio
• transcript_word_count: Count of transcript words
• transcript_char_count: Count of transcript characters
• audio: Audio file (16kHz mono WAV) with array and sampling rate
• transcript: Normalized lowercase transcript (no punctuation)

Dataset Visualization Examples

Below are example visualizations you can generate from the dataset using Python. These help explore audio characteristics and transcript distributions.

1. Audio Waveform and Mel Spectrogram (Single Sample)

from datasets import load_dataset
import librosa.display
import matplotlib.pyplot as plt
import numpy as np

dataset = load_dataset("andreoosthuizen/afrikaans-30s", split="train")

sample = dataset[0]
audio = sample["audio"]["array"]
sr = sample["audio"]["sampling_rate"]
transcript = sample["transcript"]

print(f"Transcript: {transcript}")

# Waveform
plt.figure(figsize=(14, 5))
librosa.display.waveshow(audio, sr=sr)
plt.title("Audio Waveform")
plt.xlabel("Time (s)")
plt.ylabel("Amplitude")
plt.tight_layout()
plt.show()

# Mel Spectrogram (Whisper input representation)
S = librosa.feature.melspectrogram(y=audio, sr=sr, n_mels=128)
S_dB = librosa.power_to_db(S, ref=np.max)

plt.figure(figsize=(14, 5))
librosa.display.specshow(S_dB, x_axis='time', y_axis='mel', sr=sr)
plt.colorbar(format='%+2.0f dB')
plt.title("Mel Spectrogram")
plt.tight_layout()
plt.show()

2. Transcript Length Distribution (Histogram)

import matplotlib.pyplot as plt
from datasets import load_dataset

dataset = load_dataset("andreoosthuizen/afrikaans-30s")

lengths = [len(example["transcript"].split()) for example in dataset["train"]]

plt.figure(figsize=(10, 6))
plt.hist(lengths, bins=50, color='skyblue', edgecolor='black')
plt.title("Distribution of Transcript Word Counts")
plt.xlabel("Number of Words")
plt.ylabel("Frequency")
plt.grid(axis='y', alpha=0.75)
plt.show()

3. Word Cloud of Common Words

from wordcloud import WordCloud
import matplotlib.pyplot as plt
from datasets import load_dataset

dataset = load_dataset("andreoosthuizen/afrikaans-30s", split="train")

text = " ".join(example["transcript"] for example in dataset)

wordcloud = WordCloud(width=800, height=400, background_color='white').generate(text)

plt.figure(figsize=(12, 6))
plt.imshow(wordcloud, interpolation='bilinear')
plt.axis("off")
plt.title("Word Cloud of Common Afrikaans Terms")
plt.show()

These examples demonstrate typical audio quality, spectrogram features, and linguistic patterns in the dataset.

Dataset Creation

Source Data

• Afrikaans audio of the NG Kranztkloof and NG Westville communities
• Raw audio: Variable quality, resampled to 16kHz mono

Considerations

• Bias: Content skewed toward South African religious videos (sermons, announcements)
• Noise: Real-world audio (background noise, music, overlaps)
• Ethics: Derived from public sermons; no personal data; for research only
• Limitations: Normalization removes punctuation

Additional Information

Licensing

Released under CC-BY-4.0 for research/non-commercial use.

Citation

If using this dataset, please cite:

@dataset{afrikaans_30s_2026,
  author = {André Oosthuizen},
  title = {Afrikaans Speech Dataset for Whisper Fine-Tuning},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/andreoosthuizen/afrikaans-30s}
}

Thank you for using this dataset to improve Afrikaans ASR!