Lwasinam/amharic-xtts-v2-spitchai-500h

Amharic XTTS-v2 (Adapted) by Spitch AI

Amharic TTS model adapted from XTTS-v2, trained on approximately 500 hours of data by Spitch AI (Lagos, Nigeria).

Repository Contents

• checkpoint_232000.pth (GPT checkpoint)
• dvae.pth
• mel_stats.pth
• config.json
• vocab.json
• requirements.txt
• amharic_financial_normalizer.py

Quick Setup

pip install -r requirements.txt

Inference (Python)

import pandas as pd
if not hasattr(pd.DataFrame, 'map'):
    pd.DataFrame.map = pd.DataFrame.applymap

import torch
import torchaudio
import epitran
from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import Xtts
from amharic_financial_normalizer import normalize_amharic_financial_text

CHECKPOINT_PATH = "checkpoint_232000.pth"
CONFIG_PATH = "config.json"
VOCAB_PATH = "vocab.json"
DVAE_PATH = "dvae.pth"
MEL_STATS_PATH = "mel_stats.pth"

config = XttsConfig()
config.load_json(CONFIG_PATH)
config.model_args.dvae_checkpoint = DVAE_PATH
config.model_args.mel_norm_file = MEL_STATS_PATH
config.model_args.vocab_file = VOCAB_PATH

model = Xtts.init_from_config(config)
model.load_checkpoint(
    config,
    checkpoint_path=CHECKPOINT_PATH,
    vocab_path=VOCAB_PATH,
    use_deepspeed=False,
)

if torch.cuda.is_available():
    model.cuda()

epi_am = epitran.Epitran("amh-Ethi-pp")


def preprocess_text(text: str, lang: str = "am") -> str:
    if lang == "am":
        # 1) Normalize numbers and money expressions first.
        text = normalize_amharic_financial_text(text)
        # 2) Transliterate Ethiopic script for model robustness.
        text = epi_am.transliterate(text)
    return text


ref_audio = "reference.wav"  # 3-10 seconds, clean speech
text_am = "ወደ 77777 ብር 250.50 ያስገቡ"
processed_text = preprocess_text(text_am, "am")

gpt_cond_latent, speaker_embedding = model.get_conditioning_latents(audio_path=[ref_audio])
out = model.inference(
    text=processed_text,
    language="am",
    gpt_cond_latent=gpt_cond_latent,
    speaker_embedding=speaker_embedding,
    temperature=0.5,
    length_penalty=1.0,
    repetition_penalty=2.0,
    top_k=50,
    top_p=0.8,
    enable_text_splitting=True,
)

torchaudio.save("output_amharic.wav", torch.tensor(out["wav"]).unsqueeze(0), 24000)
print("Saved output_amharic.wav")

Notes

• normalize_amharic_financial_text(...) handles amounts, decimals, phone-like IDs, and short-codes for TTS-friendly Amharic speech.
• Best results come from high-quality reference audio.
• Output sample rate is 24000 Hz.