poto-tts

Ghanaian English speech synthesis on sherpa-onnx. A 104,623-word Ghanaian lexicon is compiled into an espeak-ng dictionary, so names, places and titles come out right from plain text.

                stock espeak                  poto-tts
Kwabena         kwˈeɪbnə                      kwɑːbˈɪnɑː
Achimota        ɐtʃɪmˈoʊɾə                    ɑːtʃimˈoɾɑː
Okuapenhene     ˈoʊkjuːˌeɪpənhˌiːn            okwɑːpɛnhˈɛnɛ
Nyankpani       nˌaɪɐŋkpˈɑːni                 njɑːŋkpˈɑːni
Gyasi           dʒaɪʲˈɑːsi                    dʒˈɑːsi

Ordinary English is untouched: yesterday keeps its own stress. Compiling the lexicon in, rather than looking it up at runtime, means callers ship no lexicon file and words the lexicon never had are still spoken — espeak's own rules handle them.

pip install poto-tts
poto-tts "Kwabena went to Achimota" -o out.wav

from poto_tts import load

tts = load("kokoro")            # downloads the voice on first use
tts.save("The Okuapenhene met Nana Bawumia", "out.wav")

Backends

Both read the same espeak-ng-data, so either pronounces a Ghanaian name the same way; what differs is the voice and the licence. The piper corpus is broadcast recordings whose speakers did not consent to being modelled — research and non-commercial use only. For a commercial Ghanaian voice use kokoro, or train piper on recordings you hold rights to (tools/ takes any dataset).

tts = load("piper", voice="gh_00")
print(tts.licence, tts.commercial_use)

Models

Voices live on the Hub and are fetched through huggingface_hub, so downloads are counted:

Each repo's config.json names its own files, speakers and licence, so a new voice needs no release of this library:

tts = load("kokoro", repo_id="your-org/your-voice")

POTO_TTS_CACHE moves model files off ~/.cache.

Cross-platform

sherpa-onnx runs on Android, iOS, WebAssembly, C++, C, Go, C#, Java, Kotlin, Rust, Dart and Swift, across x86, arm64, arm32 and RISC-V. Pronunciation lives in a data directory, not in Python, so a voice is a file set that any of those runtimes loads directly — offline, no server:

A stock espeak-ng-data leaves a working voice that mispronounces every Ghanaian name, with no error. That directory is the deliverable.

Point your platform's sherpa-onnx TTS API at those files — Kotlin OfflineTts, Swift SherpaOnnxOfflineTts, the WASM build, or sherpa-onnx-offline-tts. Upstream docs.

REST API

pip install 'poto-tts[api]'
poto-tts serve --host 0.0.0.0 --port 8080
curl "localhost:8080/speak?text=Kwabena+went+to+Achimota" -o out.wav

Requests over 2,000 characters are refused rather than truncated (POTO_TTS_MAX_CHARS). For phone apps, ship the files instead — no network, no latency, no server.

Changing how a word is said

The dictionary will still miss your grandmother's name. Rebuilding is the supported fix: it changes the data every runtime loads, so Android and iOS get it too. Needs the espeak-ng binary and the lexicon extra:

pip install 'poto-tts[lexicon]'

# my_words.tsv — word<TAB>pronunciation
Darkoa	d a r k o a
Kufuor	=kufu'or

poto-tts dict --out my-espeak-data --extra my_words.tsv
POTO_TTS_ESPEAK_DATA=./my-espeak-data poto-tts "Darkoa met Kufuor" -o out.wav

Values are Ghanaian IPA, mapped by the same tables as the packaged lexicon. A leading = passes raw espeak mnemonics through, and is checked: an invalid mnemonic makes espeak silently discard the rest of an entry, so the build rejects it rather than shipping half a word. Your entries override the packaged lexicon, so the file corrects entries as well as adding them.

For a quick check without rebuilding:

poto-tts --phonemes "Nana Addo met Kwame Nkrumah"

from poto_tts import GhanaInjector
GhanaInjector(lexicon={"Darkoa": "d a r k o a"})("Darkoa came home")
# "[[dA:rk'oA:]] came home"

That is a debugging tool, not a deployment route: Python-only, and Kokoro's frontend rewrites : in its input, which corrupts length marks.

How the dictionary is built

ghana-english-g2p's Ghanaian IPA is mapped to espeak mnemonics (poto_tts/mnemonics.py) and compiled into espeak's en_extra. Entries are written only for words that are not ordinary English, checked against a 370k-word list: espeak already handles English, including its stress. An earlier version that entered every word whose pronunciation differed took 93,526 of them and moved the stress in yesterday, January and Wednesday.

The build verifies its own output — no truncated entries, and espeak's English unchanged byte for byte. The vowel decisions are in the module's comments: /a/ uses A: rather than 0, which espeak reads as a different vowel before /r/; æ merges into /a/; FACE and GOAT map to the monophthongs espeak's English rules never emit but its tables define.

Training a voice

tools/ holds the pipeline behind the piper backend, and takes any dataset:

Thresholds there are measured, and the comments record the distributions they came from — including the ones that were wrong first.

Credits

sherpa-onnx (k2-fsa) · Kokoro-82M (hexgrad, Apache-2.0) · piper1-gpl (OHF Voice) · ghana-english-g2p (GhanaNLP) · espeak-ng

MIT, except the bundled espeak-ng-data (GPL-3.0) and the piper voice (non-commercial).