Datasetpublic

michsethowusu/melo-tts-ghana-multilingual

Melo-TTS multilingual Ghana training dataset Ghanaian multilingual TTS training data for a MeloTTS fine-tune, phonemised with ghananlpcommunity/xlsr-twi-codeswitch-ipa. Languages twi-asantee — Asante Twi (AfriSpeech open-bible-speech-african) twi-akuapem — Akuapem Twi (AfriSpeech open-bible-speech-african) ewe — Ewe (AfriSpeech open-bible-speech-african) twi-en-codeswitch — Twi-English code-switched speech (ghananlpcommunity) ghana-english —… See the full description on the dataset page: https://huggingface.co/datasets/michsethowusu/melo-tts-ghana-multilingual.

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes337downloads

Dataset Card

Melo-TTS multilingual Ghana training dataset

Ghanaian multilingual TTS training data for a MeloTTS fine-tune, phonemised with ghananlpcommunity/xlsr-twi-codeswitch-ipa.

Languages

—twi-asantee — Asante Twi (AfriSpeech open-bible-speech-african)
—twi-akuapem — Akuapem Twi (AfriSpeech open-bible-speech-african)
—ewe — Ewe (AfriSpeech open-bible-speech-african)
—twi-en-codeswitch — Twi-English code-switched speech (ghananlpcommunity)
—ghana-english — Ghanaian English (ghanaopendata/ghana-english-tts-clean2)

Columns

—audio — 44.1 kHz mono audio (FLAC)
—text — source transcript
—phonemes — XLSR phoneme-recogniser output (space-separated 79-token inventory, see symbols.txt)
—ipa_reference — reference ghanag2p-uni IPA where available (code-switch rows)
—language, speaker, split, duration, source, id
—book/chapter/verse for Bible rows

Phonemisation

Every clip was phonemised with ghananlpcommunity/xlsr-twi-codeswitch-ipa (16 kHz input, CTC decode). The token inventory in symbols.txt is the MeloTTS phoneme symbol set for step-2 training.

Splits

—train — bible (train) + code-switch (train) + ghana-english sample
—validation — code-switch (validation)
—test — bible (test) + code-switch (test)

Licenses

AfriSpeech bible audio: CC BY-SA 4.0 · code-switch: Apache-2.0 · ghana-english: MIT