speech-g2p

Tier 1 · engines. No titan dependencies; one npm dependency, compromise 14.18.0 (pinned exactly). No model downloads and no network.

npm install @titan-design/speech-g2p

The problem it solves

Kokoro reads misaki phonemes, and misaki is Python: its G2P runs spaCy's tokenizer and tagger, then a lexicon with tag-keyed readings, stress rules and number reading. Without Python, a TypeScript product could not produce Kokoro's input, honour [word](/phonemes/) overrides, or dump the per-token readings that pronunciation checks need.

This package ports misaki 0.9.4's US English G2P and spaCy 3.8's English tokenizer to TypeScript, and tags with compromise through a Penn Treebank adapter. misakiEnglish() is the one primitive: text in, misaki tokens and the joined phoneme string out.

When to reach for it

You need phonemes for English text: Kokoro input through speech-tts (ttsTokens), a heteronym check through speech-vet (a G2pToken[] is a PhonemeToken[]), or IPA for a caption or a listener (misakiToIpa). Synthesis itself is speech-tts; deciding whether a reading is the declared one is speech-vet.

Example

Verified against the 0.0.0 workspace build.

import { misakiEnglish, misakiToIpa, ttsTokens } from "@titan-design/speech-g2p";

const g2p = misakiEnglish();
const { phonemes, tokens } = g2p.phonemize("We live here.");
phonemes;                    // "wˌi lˈɪv hˈɪɹ."
tokens[1];                   // { text: "live", tag: "VBP", phonemes: "lˈɪv", whitespace: " ", rating: 4 }
misakiToIpa(phonemes);       // IPA, stress kept
ttsTokens(tokens);           // speech-tts TtsToken[]: null phonemes become "" (Kokoro's unk)
g2p.version;                 // "misaki-0.9.4+<16 hex>", a cache key

What it deliberately does not do

• No espeak. A word outside misaki's lexicon is returned unread unless the caller passes fallback; misaki itself would ask espeak (GPL) for it.
• No British English. Only the US lexicon is vendored.
• No chunking at Kokoro's 510-phoneme limit and no word timings; those belong to the Kokoro provider.
• spacyTokenizer and compromiseTagger are not exported. misakiEnglish uses them by default; pass your own tokenizer or tagger to replace them.

Gotchas

• Without a fallback, out-of-lexicon words read as phonemes: null, and a compound such as file_name.txt or usr/local/bin keeps only its known parts. Kokoro skips the missing sounds silently, so warn on phonemes === null before synthesis. On a 2,839-unit test corpus of mostly fiction, this costs about 2 % of word tokens.
• Tagging is compromise's, not spaCy's. On the CI golden, 99.3 % of non-heteronym word tokens match misaki with spaCy exactly, stress included. Most of the remainder is stress on that and in. Heteronyms (read, use, live, ...) are where the tagger matters most: declare them with speech-vet.
• compromise is pinned exactly because its tags move between releases; a bump changes version and needs the parity tests rerun.
• rating is mostly null: misaki records it on only some paths. Do not rely on it.
• Construction reads about 6 MB of lexicon JSON from data/; build one misakiEnglish() and reuse it.

Where it came from

New (TP-3027). It replaces the Python misaki process in the speech pipeline. It reaches 99.42 % stress-strict parity with misaki and spaCy on the E4 spike's 2,839-unit corpus, and the tokenizer matches spaCy on every unit. The port follows misaki 0.9.4 (Apache-2.0) and spaCy 3.8.16 (MIT); see the package NOTICE.