HiTZ/wikipedia_basque_ipa

Basque Wikipedia Phonemized Corpus (Text + IPA phonemes)

Dataset Description

A large-scale paired corpus derived from the Basque Wikipedia dump. Each row contains both the original plain text and its IPA phoneme transcription, at paragraph level. Stressed vowels use the apostrophe convention (e.g. 'a, 'e, 'i, 'o, 'u) and affricates are kept as multicharacter sequences (e.g. tʃ, tʂ, ts).

This dataset is intended for training text-to-speech (TTS) and grapheme-to-phoneme (G2P) models for Basque.

---

Dataset Statistics

Split	Rows
samples	1,672,981

---

Dataset Structure

Fields

• text (string): The Basque Wikipedia plain text, at paragraph level.
• phonemes (string): The corresponding IPA phoneme transcription. Words are space-separated; punctuation is attached directly to the preceding word (e.g. astr'onomia. not astr'onomia .).

Example

from datasets import load_dataset

ds = load_dataset("HiTZ/wikipedia_basque_ipa", split="train")
print(ds[0]["text"])
# → "Historiako lehenengo zientzia izan da astronomia. Zibilizazio eta kultura guztiek..."
print(ds[0]["phonemes"])
# → "'istoɾiako le'enenɡo ʂi'entʂia iʂ'an da astr'onomia. ʂiβ'iliʂaʂio eta kult'uɾa..."

---

IPA Symbol Conventions

This dataset uses standard IPA symbols.

---

Data Processing Pipeline

Data processing steps from raw data extracted from Wikipedia to the phonemized utterances.

Step 1 — Wikipedia Extraction (WikiExtractor.py)

Raw text was extracted from the Basque Wikipedia XML dump using a customized version of WikiExtractor. The extractor:

• Strips MediaWiki markup (templates, tables, infoboxes)
• Expands wiki links to their anchor text
• Removes HTML tags, comments, and special elements (<ref>, <math>, etc.)
• Replaces <math> blocks with formula_N placeholders
• Outputs plain paragraphs, one per line

Step 2 — Cleaning

The extracted text was further cleaned with the following filters and transformations:

Sentence-level filters (removal):

• Sentences shorter than 100 characters
• Sentences containing double quotes (")
• Sentences containing chess notation (e4, c4)
• Sentences containing HTML-like symbols (<, >)
• Sentences containing the string formula kimikoa (Basque for "chemical formula")
• Sentences containing formula_N placeholders (Wikipedia math markup artifacts)
• Sentences containing | (MediaWiki pipe/table syntax artifacts)
• Sentences consisting only of digits and punctuation

Text normalization transformations:

• "K. a." → "K.a." and "K. o." → "K.o." (Basque grammatical abbreviations)
• Dots after single uppercase initials removed (e.g. A. → A)
• Content inside parentheses/brackets removed
• All bracket characters []<>{}() removed
• Hyphens between words removed (e.g. behaketa-saioa → behaketa saioa), preserving numeric ranges
• Quotes and apostrophes (", ', ", ", ', ') removed
• URLs and email addresses removed
• Non-printable characters removed
• Whitespace normalized (multiple spaces → single space)
• Consecutive or redundant punctuation cleaned up (e.g. ,, → ,, .. → .)
• Trailing punctuation normalized to a single .

Step 3 — Normalization + Phonemization

Each cleaned sentence was first normalized and then phonemized using ahoNT, a Basque text processing and phonemization tool developed at HiTZ Zentroa / AhoLab.

Normalization handles:

• Number expansion (e.g. 42 → spoken Basque word form)
• Abbreviation resolution
• Other text-to-speech pre-processing rules specific to Basque

Phonemization then:

• Converts each normalized word to its phoneme sequence
• Outputs IPA symbols with stress markers and multicharacter affricates preserved
• Attaches punctuation marks to the preceding word

---

Usage

from datasets import load_dataset

ds = load_dataset("HiTZ/wikipedia_basque_ipa", split="train")

for example in ds.select(range(5)):
    print(example["text"])
    print(example["phonemes"])
    print()

Split phonemes into individual word tokens:

for example in ds.select(range(5)):
    tokens = example["phonemes"].split()
    print(tokens)
    # e.g. ["'istoɾiako", "le'enenɡo", "ʂi'entʂia", "iʂ'an", "da", "astr'onomia.", ...]

Use as a G2P training corpus:

for example in ds.select(range(5)):
    words = example["text"].split()
    phonemes = example["phonemes"].split()
    # Note: words and phoneme tokens are aligned one-to-one
    # (punctuation is attached to the preceding phoneme token)

---

License

The dataset is derived from Basque Wikipedia, which is released under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.

---

Citation

If you use this dataset, please cite the Basque Wikipedia and acknowledge the phonemization pipeline developed at AhoLab (University of the Basque Country).

Related Resources

• ahoNT — Basque text normalization and phonemization tool
• WikiExtractor - Python script that extracts and cleans text from a Wikipedia database backup dump