ahotts-g2p

Pure-Python, version-aware grapheme-to-phoneme (G2P) for Basque (euskara) and Spanish, faithfully reproducing the AhoTTS text-to-speech front-end.

ahotts-g2p turns text into the single-char IPA training string used by StyleTTS2/VITS-style models, matching the AhoTTS engines that phonemized the public HiTZ Basque voices. The g2p pipeline itself is stdlib-only: no C build, no shared libraries. The only runtime dependency is scriptconv, used to render output in notations other than the native alphabet (see Output alphabets).

Install

pip install ahotts-g2p

From source:

pip install -e .[test]

Quick start

from ahotts_g2p import phonemize

phonemize("Bai.")                                    # 'bAj .'
phonemize("Ez, horrek ez du balio!")                 # 'Eʂ , Orek eʂ tU βalIo !'
phonemize("Kaixo mundua", lang="eu", version="classic")   # 'kajʃO mundUa'
phonemize("Hola mundo.", lang="es", version="classic")    # 'Ola mUndo'

# Northern (Iparralde / Iparrahotsa) Basque dialect
phonemize("hori horrek", lang="eu", dialect="northern")   # 'hOɾi hoʁEk'

CLI:

python -m ahotts_g2p "Kaixo mundua"
cat sentences.txt | python -m ahotts_g2p

Also exported: SAMPA_TO_IPA, the ordered SAMPA -> IPA mapping table.

Output alphabets

phonemize() takes an alphabet argument. The default, "native", is the AhoTTS single-char training string above -- unchanged from previous releases.

from ahotts_g2p import phonemize, SUPPORTED_ALPHABETS

print(SUPPORTED_ALPHABETS)
# ('native', 'ipa', 'x-sampa', 'arpa', 'lexique', 'kirshenbaum', 'cotovia', 'rfe', 'halabi')

phonemize("Bai.", alphabet="ipa")                                        # 'bˈaj .'
phonemize("Kaixo mundua", version="classic", alphabet="x-sampa")         # 'kajS"o mund"ua'
phonemize("Hola mundo.", lang="es", version="classic", alphabet="ipa")   # 'ˈola mˈundo'

CLI:

python -m ahotts_g2p --alphabet x-sampa "Kaixo mundua"

alphabet="ipa" expands the native string back to plain IPA: it reverses the folding phonemize() applies for the single-char training format (affricates C/V/P, stressed vowels I/E/A/O/U, the Spanish aspirated stops H/K/T), and corrects the ASCII apostrophe the AhoTTS dictionaries use internally for stress to the real IPA primary-stress mark, ˈ (U+02C8). Every other native character is already IPA, per SAMPA_TO_IPA, so most phones need no change at all. Every other alphabet value is a scriptconv phonetic notation, reached by converting that IPA string with scriptconv.notation.convert.

Punctuation tokens (., ,, !, ?, ;, : -- surfaced by the modern engine as their own space-separated token) pass through unchanged in every alphabet: they are not phones, and no phonetic notation covers them.

A phone with no mapping in the requested notation raises ValueError -- it is never silently dropped. The X-SAMPA trill r and plain h gaps present in older scriptconv releases are both fixed as of scriptconv 0.0.4a16 (the alveolar trill and the voiceless glottal fricative now both encode fine). ARPABET remains a real, permanent gap for several Basque/Spanish phones -- it is scoped to English phonology and will never grow entries for them:

• ʂ (retroflex fricative), e.g. Basque /z/ in "Ez", raises when converted to alphabet="arpa".
• β (voiced bilabial approximant/fricative), intervocalic Basque/ Spanish b/v (e.g. in "ibili"), raises the same way.

Versions

AhoTTS has a real engine lineage. Different public voices were phonemized by different generations, with visibly different output, so the API takes a version (classic/modern). The default is modern.

(pyAhoTTS builds the classic engine from ekaitz-zarraga/AhoTTS, a packaging fork of aholab/AhoTTS with build/portability changes only -- no algorithmic difference.)

Full detail in docs/versions.md.

Dialects

Basque has a Northern (Iparralde) variety with its own AhoTTS engine, AhoTTS_Iparrahotsa. It is exposed as a dialect (dialect="northern", default "standard"), independent of version:

phonemize("Euskara Euskal Herriko hizkuntza da.", lang="eu", dialect="northern")
# 'Ewʂkaɾa ewʂkAl heʁIko hiskUnVa ðA'

The Northern dialect pronounces /h/, has the French vowel ü -> /y/, a uvular rhotic /ʁ/, a remapped sibilant system (s -> ʂ, z -> s, ts -> tʂ), and j/dd -> /ɟ/. It is a faithful port of the AhoTTS_Iparrahotsa fork. Full detail in docs/dialects.md.

Accuracy

Correctness is parity with the AhoTTS reference engines, measured per version on held-out corpora (positional word match):

The Northern Basque dialect reaches 99.61% word parity (418/430 exact lines) against the AhoTTS_Iparrahotsa binary; see docs/dialects.md.

The held-out corpora ship as test fixtures, so the figures reproduce with no binaries: pytest tests/test_oracle.py. See docs/accuracy.md.

Pipeline

text -> normalize -> g2p -> syllabify -> stress -> SAMPA -> IPA -> single-char

Numbers, ordinals and roman numerals are expanded to the target-language number words; punctuation is preserved as separate tokens (modern) or dropped (classic). Per- word lexical stress and the phonetic-exception rules are driven by the decoded dictionary flags. See docs/architecture.md.

Supported languages

• Basque (eu) -- full linguistic pipeline with HDIC-dictionary POS tagging and accentual-group stress.
• Spanish (es) -- dictionary-free g2p and stress.

Where it fits

License

GPL-3.0-or-later, matching upstream AhoTTS. This is a derivative of the GPL AhoTTS linguistic rules, so it is distributed under the same licence. The AhoTTS algorithms and dictionaries are credited to Aholab (UPV/EHU). See docs/licensing.md.