README.md

metadata

license: mit
language:
  - en
  - de
tags:
  - g2p
  - grapheme-to-phoneme
  - ipa
  - text-to-speech
pretty_name: Crane G2P Test Fixtures

G2P Test Fixtures

Held-out word-to-IPA test sets for evaluating grapheme-to-phoneme accuracy, used by Crane's test suite and benchmarks. These are static test fixtures, not runtime model assets.

Files

File	Language	Entries	Description
g2p/en_us/test.tsv	English (US)	5000	Held-out word-IPA pairs for CER benchmarking
g2p/de_de/test.tsv	German	5000	Held-out word-IPA pairs for CER benchmarking
g2p/kokoro_vocab.json	n/a	114	Kokoro phoneme vocabulary (char → token ID) used by the normalizer correctness benchmark
g2p/en_us/kokoro_normalizer_ref.tsv	English (US)	71	Raw IPA → expected (en_us, Kokoro) normalized IPA + phoneme IDs, for IpaNormalizer regression testing
g2p/de_de/kokoro_normalizer_ref.tsv	German	32	Raw IPA → expected (de, Kokoro) normalized IPA + phoneme IDs, for IpaNormalizer regression testing

Format

g2p/en_us/test.tsv and g2p/de_de/test.tsv: tab-separated, two columns, UTF-8, no header row, sorted alphabetically:

word<TAB>IPA

Example:

abandons	əbˈændənz

g2p/en_us/kokoro_normalizer_ref.tsv and g2p/de_de/kokoro_normalizer_ref.tsv: tab-separated, three columns, UTF-8, no header row:

raw_ipa<TAB>expected_normalized_ipa<TAB>expected_phoneme_ids

expected_phoneme_ids is a comma-separated list of i64 token IDs, with a 0 (pad/BOS) prepended and appended, matching Moonshine's phoneme_str_to_input_ids(). Example:

tˈiːtʃɚ	tˈiːʧəɹ	0,62,156,51,158,133,83,123,0

g2p/kokoro_vocab.json: a flat JSON object mapping each phoneme character to its integer token ID, e.g. {"a": 43, "ˈ": 156, ...}.

Provenance

g2p/en_us/test.tsv is a 5000-entry random sample (seed 42, no replacement) from Moonshine Voice's (https://github.com/moonshine-ai/moonshine, MIT license) English IPA lexicon, which is itself derived from the CMU Pronouncing Dictionary (https://github.com/cmusphinx/cmudict, BSD-style license). These entries are held out and never seen by the G2P engine during lexicon construction.

g2p/de_de/test.tsv is a 5000-entry random sample (seed 42, no replacement, one pronunciation per word) from Moonshine Voice's German IPA lexicon (MIT-licensed multilingual IPA dictionary data), deduplicated to unique words before sampling since the source lexicon has multiple pronunciation rows for some words. These entries are held out and never seen by the G2P engine during lexicon construction.

g2p/kokoro_vocab.json is copied verbatim from Moonshine Voice's core/moonshine-tts/data/kokoro/config.json vocab object (114 entries; the HuggingFace onnx-community/Kokoro-82M-v1.0-ONNX export used by Crane's kokoro_tts module has the same 114 entries plus one extra, $, which does not affect this normalizer's behavior).

g2p/en_us/kokoro_normalizer_ref.tsv combines the first 50 lines of Moonshine's own golden English G2P output (core/moonshine-tts/tests/data/en_us/rule_g2p_wiki_100.txt) with 21 hand-crafted edge cases (empty/whitespace-only input, each of the 12 diphthong/affricate/rhotic replacements in isolation, unknown-codepoint dropping, non-ASCII whitespace collapsing). Expected output was generated by an independent Python reimplementation of Moonshine's normalize_ipa_to_kokoro() and phoneme_str_to_input_ids() (moonshine-tts.cpp), not by running the C++ itself — the 12 replacement pairs and the vocab-filter/whitespace-collapse logic are simple enough that this port only needs str.replace, str.strip, and Python's unicodedata.normalize("NFC", ...).

g2p/de_de/kokoro_normalizer_ref.tsv combines 6 realistic German lexicon entries sampled from g2p/de_de/test.tsv (single words plus one 2-word combination, chosen to exercise both plain lexicon passthrough and the affricate/geminate/ʏ replacements together with dropped combining diacritics) with 26 hand-crafted edge cases (each DE_EXTRA_KOKORO_REPLACEMENTS pair in isolation, shared affricate/diphthong passthrough under "de", German phonemes with no replacement needed, syllabic-consonant and non-syllabic-offglide diacritic dropping, unknown-codepoint dropping, and whitespace handling). Expected output was hand-derived by applying SHARED_KOKORO_REPLACEMENTS/DE_EXTRA_KOKORO_REPLACEMENTS and the vocab-filter/whitespace-collapse algorithm against g2p/kokoro_vocab.json directly, then cross-checked by running the real IpaNormalizer and confirming an exact match — not authored by trusting normalize()'s own output as ground truth.

License

Both upstream sources use permissive, MIT-compatible licenses:

• Moonshine Voice: MIT license.
• CMU Pronouncing Dictionary: BSD-style license, Copyright (C) 1993-2015 Carnegie Mellon University. Full text: https://github.com/cmusphinx/cmudict/blob/master/LICENSE