README.md

metadata

license: cc0-1.0
task_categories:
  - text-to-speech
  - automatic-speech-recognition
language:
  - en
tags:
  - australian
  - australian-english
  - australian-accent
  - accent
  - english-accent
  - speech
  - voice
  - audio
  - tts
  - text-to-speech
  - asr
  - speech-recognition
  - voice-cloning
  - librivox
  - audiobook
  - public-domain
pretty_name: Australian English Speech (Australian accent audio dataset, LibriVox)
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Australian English Speech

An Australian accent speech dataset: 110 hours of Australian English audio with transcripts, from the Australian titles in the LibriVox catalogue, cut into short clips for text-to-speech, voice cloning and speech recognition.

Open English speech corpora are overwhelmingly American. This one is Australian, public domain, and free of any redistribution or model-release restrictions.

LibriVox recordings are released into the public domain, and the books themselves are public domain, so this dataset is unencumbered: train on it, redistribute it, ship models built from it.

What is in it

clips	61,662
audio	110.2 hours
readers	214
books	37
format	mono FLAC, 24 kHz
clip length	3-15 s

Each row:

field	meaning
audio	the clip, 24 kHz mono
text	transcript
reader	LibriVox narrator's display name
reader_id	LibriVox reader id
book	book title
author	author(s)
book_id	LibriVox book id
section	section number within the book
duration	clip length in seconds

from datasets import load_dataset

ds = load_dataset("ablmontazer/australian-english-speech", split="train")
print(ds[0]["text"], ds[0]["audio"]["sampling_rate"])

reader is a stable speaker label: filter on it to get single-speaker data.

How it was built

1. Australian titles taken from LibriVox's own catalogue search.
2. Each section downloaded from archive.org.
3. Gentle denoising with DeepFilterNet 3 (atten_lim_db=6). LibriVox audio is already clean, so this only shaves the hiss rather than reshaping the voice.
4. Silero VAD segmentation into 3-15 s clips that start and end on silence.
5. Transcription with faster-whisper large-v3-turbo.
6. Clips whose transcript is LibriVox boilerplate (the spoken intro and outro that top and tail every section) are dropped.

Things to know before you train on it

• Transcripts are machine-generated. Whisper is good but not perfect, and punctuation and capitalisation are inconsistent between clips. For anything where exact text matters, re-check against the public-domain source texts.
• Accent is not verified per reader. These are the Australian titles in the catalogue; LibriVox volunteers are international, so a minority of narrators reading Australian books are not themselves Australian. Filter by reader if you need a guaranteed accent.
• Per-reader volume is very uneven, from a few minutes to over ten hours. Balance or cap by reader if that matters for your training.
• Audio quality varies with the narrator's home recording setup, and some of the older recordings are effectively band-limited.

Credits

Every recording here was made by a LibriVox volunteer who donated it to the public domain. The narrators, and the books they read, are listed in CREDITS.md.