ghanaopenai/ghana-voice-base

Neu-TTS Ghana Foundation Base Model (Zero-Ref IPA)

ghanaopenai/neu-tts-ghana-base is the multilingual text-to-speech foundation base model for Ghanaian languages, operating in direct zero-reference mode (no reference audio prompt required).

Trained to full convergence using Joint Loss + CER Plateau detection across 12,695 clean, digit-free speech clips selected from AfriSpeech Selector.

---

Key Features

1. Direct Zero-Reference TTS: Direct text-to-speech generation without needing reference audio prompts or reference acoustic tokens.
2. Word-Separated IPA Conditioning:
   • Inside each word, characters have NO spaces (akʷaaba, oɲankʰopʰɔn).
   • Words are separated by single spaces.
   • Punctuation (, . ? ! ; :) is preserved for natural pauses and rhythm.
   • Zero language tags (pure phonetic conditioning).
3. 13 Ghanaian Languages Supported:
   • Asante Twi (twi)
   • Akuapem Twi (twi)
   • Fante (fat)
   • Ga (gaa)
   • Ewe (ewe)
   • Dagbani (dag)
   • Dagaare (dga)
   • Frafra / Farefare (gur)
   • Nzema (nzi)
   • Ahanta (aha)
   • Sehwi (sfw)
   • Dangme (ada)
   • Ghanaian English (eng)
4. NeuCodec Audio Decoding: Generates discrete acoustic tokens at 50 Hz frame rate, decoded with neuphonic/neucodec into 24 kHz high-fidelity audio.

---

Quickstart Inference

import torch
import soundfile as sf
import re
from transformers import AutoModelForCausalLM, AutoTokenizer
from africa_g2p import AfricaPipeline
from neucodec import NeuCodec

device = "cuda" if torch.cuda.is_available() else "cpu"
model_id = "ghanaopenai/neu-tts-ghana-base"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map=device)
codec = NeuCodec.from_pretrained("neuphonic/neucodec").to(device).eval()

# 1. Phonemize text to word-separated IPA
pipe = AfricaPipeline(lang="twi", output="ipa")
ipa_text = pipe.run("Akwaaba, yɛbɔ mpae sɛ Onyankopɔn bɛhyira wo.", sep="").strip()

# 2. Format zero-reference prompt
prompt = f"<|TEXT_PROMPT_START|>{ipa_text}<|TEXT_PROMPT_END|><|SPEECH_GENERATION_START|>"
input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device)

# 3. Generate speech tokens
speech_end_id = tokenizer.convert_tokens_to_ids("<|SPEECH_GENERATION_END|>")
out = model.generate(
    input_ids,
    eos_token_id=speech_end_id,
    max_new_tokens=400,
    do_sample=True,
    temperature=0.7,
    top_p=0.8,
    repetition_penalty=1.05
)

# 4. Decode with NeuCodec
gen_str = tokenizer.decode(out[0, input_ids.shape[1]:], add_special_tokens=False)
speech_ids = [int(n) for n in re.findall(r"<\|speech_(\d+)\|>", gen_str)]
sp_t = torch.tensor(speech_ids, device=device).unsqueeze(0).unsqueeze(0)
wav = codec.decode_code(sp_t).squeeze().cpu().float().numpy()

sf.write("output.wav", wav, 24000)
print("Saved output.wav!")

---

Training Framework

Fine-tuned and exported using GhanaOpenAI/ghana-voice-trainer.