espyak

A pure-Python reimplementation of espeak-ng's grapheme-to-phoneme (G2P) front-end. Text → phonemes only: no synthesis, no audio, no C extension, no runtime dependencies.

Reproduces the espeak-ng binary (pinned 1.52.0) byte-for-byte. The full-headword parity audit (every *_list word, force_compat mode) is 99.71% across 105 languages — 59 at 100% — and 99.69% across 22 sub-dialect variants (pt-br, en-us, es-419, …). 117 languages bundled. The remaining fails are catalogued in docs/remaining-gaps.md.

from espyak import G2P

g2p = G2P("en")
g2p.phonemize("hello world")              # 'həlˈəʊ wˈɜːld'   (IPA)
g2p.phonemize("hello world", ipa=False)   # "h@l'oU w'3:ld"   (Kirshenbaum / -x)

G2P("es").phonemize("buenos días")        # 'bwˈenos dˈias'
G2P("de").phonemize("straße")             # 'ʃtɾˈɑːsə'
G2P("ru").phonemize("привет")             # 'prʲivʲˈet'

Why

espyak gives projects espeak-ng's phonemes without the native dependency: nothing to shell out to, no C-extension to build, and the rules are readable and patchable in Python. It drops in as a backend for phoonnx.

Install

pip install -e .          # from a clone (the espeak-ng source data is bundled, ~44 MB)
# or:  uv pip install -e .

Python ≥ 3.9. The espeak-ng dictsource/, phsource/, and lang/ data are bundled under espyak/data/ at the pinned 1.52.0 tag, so nothing is needed system-wide.

Usage

Python

from espyak import G2P

g2p = G2P("en")                      # one translator per language — construct once, reuse

g2p.phonemize("read")                # 'ɹˈiːd'
g2p.phonemize("2024 dogs")           # numbers expand to words, then phonemes

g2p.phonemize("cat", ipa=True)       # 'kˈat'        — Unicode IPA (default)
g2p.phonemize("cat", ipa=False)      # "k'at"        — Kirshenbaum ASCII (espeak -x)
g2p.phonemize("cat", separator="_")  # 'k_ˈa_t'      — separate phonemes
g2p.phonemize("cat", tie="͡")         # tie multi-char phoneme names

Dialects / sub-dialect variants

A variant code (pt-br, en-us, es-419, fr-be, …) is loaded the same way — pass it to G2P. espeak-ng models a sub-dialect as a small voice file that layers over the shared base language (its dictionary, rules and translator config), overriding only the phoneme table, the dictionary conditionals, and a few post-translation phoneme substitutions:

G2P("pt-br").phonemize("dia")        # 'dʒˈiæ'   — Brazilian di palatalization
G2P("pt").phonemize("dia")           # 'dˈiɐ'    — base (European) Portuguese
G2P("en-us").phonemize("better")     # 'bˈɛɾɚ'   — rhotic, the en-us replace table
G2P("es-419").phonemize("cielo")     # 'sjˈelo'  — seseo (θ→s); base es → 'θjˈelo'

The 22 supported variants (codes are case-insensitive):

Command line

espyak -v en "hello world"           # həlˈəʊ wˈɜːld
espyak -v es "díganme"               # dˈiɣanme
espyak -v fr -x "bonjour"            # bO~Z'ur    (Kirshenbaum)
espyak -v de --sep _ "haus"          # h_ˈaʊ_s
echo "привет" | espyak -v ru -       # read from stdin

Output formats

G2P(lang).phonemize(text, ipa=True, tie=None, separator=None) is the whole surface; see docs/usage.md for details and render() (raw phoneme-string rendering).

How it works

espyak parses espeak-ng's own source data at load time and replays its pipeline in Python:

text → dictionary _list lookup → prefix/suffix retranslation → letter-to-sound rules
     → SetWordStress → phoneme programs (ChangePhoneme/InsertPhoneme) → render (IPA / -x)

Fidelity is inherited from the bundled data; the matcher, stress, number, and phoneme-program logic are re-implemented to match the binary, espeak-ng's quirks included. docs/architecture.md has the module map and pipeline.

Verification

pytest -q                            # unit + fixture tests
python test/sweep.py 25              # per-language _list-headword sweep vs the oracle
python test/corpus_sweep.py          # real-sentence corpus vs the oracle
python test/parity_audit.py --cap 2000              # full-headword parity, every language
python test/parity_audit.py --variants --cap 2000   # the same for the 22 sub-dialect variants

The reference ("oracle") is a pinned espeak-ng 1.52.0 build, used only to generate expected outputs — espyak never calls it at runtime. Every dictionary *_list headword is a free test case. parity_audit.py tests every headword in force_compat mode (the edge cases — single accented letters, abbreviations, codepoint names — are where parity breaks) and writes a per-language table plus a JSONL of every mismatch; --variants audits each dialect's base headwords through its voice layer. test/report.md holds the per-language sweep pass rate.

Coverage

Full-headword parity is 99.71% (105 languages, --cap 2000) and 99.69% across the 22 variants, with 59 base languages at 100%. The ~136 remaining base fails are concentrated in phoneme-level language switches, letter-name/abbreviation spelling (LookupDictList), unported numbers.c branches, and a small irreducible floor (formant-synthesis allophones, the ru a/ɑ reduction, oracle self-inconsistencies). They are catalogued, classified, and prioritized in docs/remaining-gaps.md. The default engine (force_compat=False) is the linguistically-correct G2P and may deviate from espeak-ng only at the entries in docs/divergences.md.

Project layout

espyak/            the engine (one module per espeak-ng translation unit)
  api.py           public G2P entry point
  dictionary.py    MatchRule / TranslateRules / SetWordStress / LookupDict2
  rule_compiler.py compiledict.c — rule byte encoding + groups
  phoneme_tab.py   phsource loader; phoneme_program.py — ChangePhoneme/InsertPhoneme
  language_data.py per-language translator config (tr_languages.c + voice files)
  voice.py         sub-dialect VARIANT loader (voices.c LoadVoice) — pt-br/en-us/es-419/…
  numbers.py       TranslateNumber + ordinals/fractions
  render.py        phoneme list → IPA / Kirshenbaum / stress / tie / separator
  data/            bundled espeak-ng dictsource/ phsource/ lang/ @ 1.52.0
docs/              architecture, usage, divergences, remaining-gaps, code-review/
examples/          runnable usage examples
test/              unit tests, oracle fixtures, sweep + corpus + parity_audit harnesses

Provenance

espyak is an AI-assisted port. The Python was written by an AI coding assistant that read and instrumented espeak-ng's C source; human review has been minimal. It is not an independent clean-room implementation.

License

espyak is GPL-3.0-or-later, the same as espeak-ng — from which it is derived and whose data it bundles under espyak/data/. See LICENSE.