200 h Ghanaian English, targets from ghana-english-g2p
Inventory
188 units: the Ghanaian 172 plus 16 English-only symbols
Why this model exists
The 42-language Ghanaian recogniser transcribes English into its own convention: serve the
comes out s ɛ f d ɛ, and symbols like ð, iː, ɜː barely appear. Measured against
ghana-english-g2p that is 76% UER, with 8.3% of units outside its inventory.
For TTS that is fatal. Training text can come from an ASR, but at synthesis time there is no
ASR — only a g2p. Labels a g2p cannot reproduce make the English portion of a voice
undriveable. So this model was finetuned on g2p targets: it still reads what the speaker
actually said, but writes it in the convention you can generate at inference —
16.49% UER against g2p, down from 76%.
The residual disagreement is wanted, not error. A model at 0% UER would merely be reproducing
g2p, adding nothing over running g2p on the text. The gap is where the recogniser hears a real
Ghanaian English realisation that canonical g2p does not predict.
Inventory
The Ghanaian model's 172 units keep their exact output ids, so each phoneme the two share
inherited a trained output row instead of starting cold — 55 of the 70 English phonemes.
Appended English-only symbols: / n̩ y æ æː ɑ ɑː ɒ ɕ ɚ ɜː ɬ ɹ ʌ θ ᵻ.
That inventory is the union of three sources: the g2p lexicon (63 phones across 104,715
words), its bundled espeak-ng fallback probed over every lexicon word (56), and the corpus
itself (67). Each contributed symbols the others missed — ɚ and ᵻ come only from espeak,
ɔ̃ only from the corpus — so any single source would have left gaps that surface at
inference as symbols the model has no id for.
Usage
Two things about this model family are easy to get wrong, and neither fails loudly:
normalise the input — per-utterance zero mean, unit variance (normalize_audio: true).
Raw waveforms still decode fluently, about 12% UER worse.
window long audio — the encoder collapses silently past ~10 s; a 30 s clip can decode to
a single unit.
