VieNeu-TTS Ghana Foundation (0.3B)

ghanaopenai/vieneu-tts-ghana-foundation is a multilingual text-to-speech foundation model for Ghanaian and West African languages, built on VieNeu-TTS-0.3B and NeuCodec.

It serves as the base model for downstream voice adaptation and single-language specialization via ghana-voice-trainer.

Model Details

Architecture: Autoregressive causal language model emitting discrete NeuCodec audio tokens (<|speech_0|> to <|speech_65535|>).
Base Model: pnnbao-ump/VieNeu-TTS-0.3B (0.3B parameters).
Training Corpus: ghanaopenai/ghana-one-voice-v2 (67,057 audio clips, 42 Ghanaian and West African languages).
Phonemization: africa-g2p universal grapheme space (shared ~45-unit phonetic inventory across all languages, without language tags).
Audio Codec: NeuCodec (neuphonic/neucodec) @ 16 kHz / 24 kHz (50 Hz frame rate).

Languages Covered (42 Subsets)

Akuapem Twi, Asante Twi, Anyin, Avatime, Bassar/Ntcham, Bimoba, Birifor (Southern), Bissa, Buli, Chumburung, Dagaare, Dagbani, Dangme, Deg, Ewe, Fante, Fulfulde, Gikyode, Gonja, Hausa, Kabiye, Kasem, Konkomba, Konni, Kusaal, Lelemi, Mampruli, Nawuri, Ninkare/Farefare, Nkonya, Ntrubo, Nzema, Paasaal, Sehwi, Sekpele, Selee, Sisaala, Siwu, Tampulma, Tem, Tuwuli, Vagla.

In-Context Voice Cloning Prompt Format

The model uses in-context voice prompting:

<|TEXT_PROMPT_START|>{ref_phonemes} {target_phonemes}<|TEXT_PROMPT_END|><|SPEECH_GENERATION_START|>{ref_codes}

The model generates continuation tokens {target_codes} matching the reference speaker's voice timbre, ending with <|SPEECH_GENERATION_END|>.

Quickstart & Finetuning

Use ghana-voice-trainer to finetune this foundation model on your own voice or language dataset:

git clone https://github.com/GhanaOpenAI/ghana-voice-trainer.git
cd ghana-voice-trainer
pip install -r requirements.txt

# Finetune on a single-speaker or custom voice dataset:
python -m ghana_voice_trainer.train \
  --base-model ghanaopenai/vieneu-tts-ghana-foundation \
  --data-dir path/to/dataset \
  --run-name my-custom-voice \
  --epochs 20

Evaluation Benchmark

Evaluated using sherpa-onnx with Meta's Omnilingual ASR 300M CTC (sherpa-onnx-omnilingual-asr-1600-languages-300M-ctc-v2):

Language	Test Character Error Rate (CER)
Dagbani	21.05%
Fante	21.88%
Dangme	24.44%
Twi	25.00%
Ewe	26.32%
Mean CER	32.06%

Acknowledgments

Trained by the GhanaOpenAI team using the NVIDIA H200 GPU infrastructure.

Downloads last month: 40

Safetensors

Model size

0.2B params

Tensor type

BF16