eryx-swahili-tts-v1

import torch
from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import Xtts
from huggingface_hub import hf_hub_download

# Download speaker embeddings
embedding_path = hf_hub_download(
    repo_id="EryxLabs/eryx-swahili-tts-v1",
    filename="swahili_speaker.pt"
)

# Load XTTS-v2 model
model_path = "path/to/xtts_v2"  # or download from coqui
config = XttsConfig()
config.load_json(f"{model_path}/config.json")
model = Xtts.init_from_config(config)
model.load_checkpoint(config, checkpoint_dir=model_path, eval=True)

# Load Swahili speaker embeddings
embeddings = torch.load(embedding_path)
gpt_cond_latent = embeddings['gpt_cond_latent']
speaker_embedding = embeddings['speaker_embedding']

# Synthesize Swahili text
# Note: Use 'en' for language since XTTS-v2 doesn't support 'sw' directly
out = model.inference(
    text="Habari yako, mimi ni msaidizi wa Kiswahili.",
    language="en",  # Swahili uses Latin script, works with English tokenizer
    gpt_cond_latent=gpt_cond_latent,
    speaker_embedding=speaker_embedding,
)

# Save audio
import torchaudio
torchaudio.save("output.wav", torch.tensor(out["wav"]).unsqueeze(0), 24000)