cstr/g2p-dicts

G2P Pronunciation Dictionaries for CrispASR

IPA pronunciation dictionaries for text-to-phoneme conversion in TTS backends (piper, kokoro, melotts). Auto-downloaded by CrispASR on first use.

Pre-generated IPA Dictionaries (primary — piper-compatible)

Pre-generated IPA transcriptions matching the phoneme inventory of piper TTS models. These are factual phonetic data — word-to-IPA mappings generated by processing vocabulary lists through a phonemization engine.

File	Language	Entries	Size	Source vocabulary
espeak_en_us.tsv	English (US)	126K	3 MB	CMUdict (BSD)
espeak_de.tsv	German	667K	23 MB	open-dict-data single words
espeak_fr.tsv	French	257K	6.6 MB	OLaPh FR (MIT)
espeak_es.tsv	Spanish	600K	18 MB	OLaPh ES (MIT)
ru_g2p_ipa.tsv	Russian	813K	36 MB	bene-ges/ru_g2p_ipa_bert_large (CC-BY-4.0)

Format: word\t/IPA/ (tab-separated, IPA in slashes, one entry per line)

Benchmark: 99.5% exact match (EN) / 100% (DE) against piper model ground truth on a 382-word test set.

Russian (ru_g2p_ipa.tsv, ru_heteronyms.txt) — CC-BY-4.0

Derived from bene-ges/ru_g2p_ipa_bert_large, licensed CC-BY-4.0. Attribution is required by that licence and is given here and in the CrispASR source beside the download URL.

Two files, both plain text:

File	Entries	Upstream file	What it is
ru_g2p_ipa.tsv	812,953	g2p_correct_vocab.txt	word → IPA, lexical stress already resolved
ru_heteronyms.txt	17,359	heteronyms.txt	words upstream judged genuinely ambiguous

Changes made to the upstream vocabulary (the only ones):

1. The stress marker was ` (U+0060, backtick) placed immediately before the stressed vowel**. It is rewritten to **ˈ (U+02C8, IPA primary stress), still immediately before the stressed vowel. U+0060 is not an IPA symbol and is absent from every TTS phoneme inventory we target; U+02C8 is in all of them. 7.57% of all symbols in the file are this marker, so leaving it as a backtick would have meant a silent 7.57% symbol loss in any consumer that filters to its own inventory.
2. IPA wrapped in /…/ and a word ipa header line added, matching the other files here.
3. 1,952 entries whose KEY was not a plain word were dropped — mostly the (по)… optional-prefix notation, which never occurs as a token in running text.

Facts a consumer needs to know about this data:

• Keys contain no ё. It is folded to е throughout, so ёлка must be looked up as елка (its IPA /ˈjɵɫkə/ does carry the ɵ). A consumer that does not fold ё→е before lookup will treat every ё word as OOV.
• Monosyllables carry no stress mark (сад → /sat/), matching espeak-ng.
• The heteronym list and the vocabulary are DISJOINT — 0 of the 17,359 heteronyms appear in ru_g2p_ipa.tsv. The heteronyms are not entries with a chosen reading; they are the words upstream REMOVED because it could not choose. So a lookup-based G2P has no reading for any of them and must fall back to letter-to-sound rules, which pick one reading without sentence context. Frequent words are in that set (все, уже, потом, чем, дорога, небо, тест). This is a property of the data, not a bug in the consumer.
• 37 distinct codepoints appear in the IPA column; after change (1) all 37 are in Zonos's phoneme inventory, so nothing is dropped by that model.

CMUdict (English ARPAbet — fallback)

File	Language	Entries	Source	License
cmudict.dict	English	126K	CMU Sphinx	BSD

Format: WORD PH1 PH2 PH3 (space-separated ARPAbet phonemes). CrispASR converts ARPAbet to IPA at runtime (76% piper match rate).

OLaPh Dictionaries (MIT — multilingual fallback)

File	Language	Entries	License
olaph_en_us.txt	English (US)	508K	MIT
olaph_de.txt	German	1.12M	MIT
olaph_fr.txt	French	259K	MIT
olaph_es.txt	Spanish	600K	MIT

Format: word\t/IPA/ (same as pre-generated dicts). OLaPh uses British English IPA conventions (ɐ for -er, ɒ for LOT); the pre-generated dicts above use American conventions matching piper models.

Source: iisys-hof/olaph — Becker et al., "OLaPh: Optimal Language Phonemizer", arXiv:2509.20086, MIT license.

Usage in CrispASR

Dictionaries are auto-downloaded to ~/.cache/crispasr/ on first TTS use:

# Default: uses pre-generated IPA dicts (piper-compatible)
crispasr --backend piper -m auto --auto-download --tts "Hello world" --tts-output out.wav

# Select dictionary source explicitly
crispasr --backend piper --g2p-dict cmudict --tts "Hello world" --tts-output out.wav
crispasr --backend piper --g2p-dict olaph --tts "Hello world" --tts-output out.wav

# Use a custom dictionary file
crispasr --backend piper --g2p-dict /path/to/my/dict.txt --tts "Hello world" --tts-output out.wav

Override per-language paths with environment variables:

• CRISPASR_CMUDICT_PATH — English CMUdict file
• CRISPASR_DE_DICT_PATH / CRISPASR_FR_DICT_PATH / CRISPASR_ES_DICT_PATH
• CRISPASR_RU_DICT_PATH — Russian dictionary file
• CRISPASR_RU_HETERONYMS_PATH — Russian heteronym list

Phonemization Cascade

CrispASR tries pronunciation sources in priority order:

1. Pre-generated IPA dict (99.5% piper match) — auto-downloaded
2. CMUdict + ARPAbet→IPA (EN, 76% match) / OLaPh (DE/FR/ES)
3. LTS rules (letter-to-sound, always available, zero dependencies)
4. espeak-ng via dlopen (loaded at runtime if installed on system)
5. espeak-ng via popen (subprocess fallback)

Attribution

• CMUdict: Carnegie Mellon University, BSD license. https://github.com/cmusphinx/cmudict
• OLaPh: Becker, M. et al. (2025). "OLaPh: Optimal Language Phonemizer." arXiv:2509.20086. MIT license. https://github.com/iisys-hof/olaph
• open-dict-data: CC-BY-SA 3.0 (Wiktionary-sourced). https://github.com/open-dict-data/ipa-dict
• Russian (ru_g2p_ipa.tsv, ru_heteronyms.txt): derived from bene-ges/ru_g2p_ipa_bert_large by bene-ges, licensed CC-BY-4.0 (https://creativecommons.org/licenses/by/4.0/). Changes: backtick stress marker rewritten to U+02C8, IPA wrapped in slashes, (по)… keys dropped. These two files remain under CC-BY-4.0.

License

Per file — this dataset is not under one licence.

File(s)	Licence
espeak_*.tsv	MIT (factual phonetic data: word→IPA mappings)
cmudict.dict	BSD (CMU Sphinx)
olaph_*.txt	MIT (iisys-hof/olaph)
ru_g2p_ipa.tsv, ru_heteronyms.txt	CC-BY-4.0 (bene-ges/ru_g2p_ipa_bert_large) — attribution required, see above

The open-dict-data vocabulary used as the German word source is CC-BY-SA 3.0.