Introduction

Hubi-TTS is a fresh open Somali speech stack on Hugging Face: ayeeyokoris/hubi-tts (created 2026-11-03 ~18:11 UTC, still being touched through 2026-11-04). The card tags language so, ships a single model.safetensors, and documents both plain TTS and reference-audio voice cloning through transformers AutoModel + trust_remote_code=True.

After a night heavy on Indic on-device ASR/TTS, this is a clean low-resource voice drop you can pull and run without a gated form.

What Shipped

  • Vendor / product: Hubi-TTS — ayeeyokoris (Hugging Face)
  • Modality: open-weight text-to-speech (+ optional voice cloning from a prompt wav)
  • Open vs closed: open weights on HF (card does not set an SPDX license field yet — treat as “card-as-published,” check the repo before commercial use)
  • Runnable today: pip install torch torchaudio transformers huggingface_hub then AutoModel.from_pretrained("ayeeyokoris/hubi-tts", trust_remote_code=True) on CUDA or CPU
  • Concrete fact: root weights are about 176 MB (model.safetensors ≈ 176,380,304 bytes); config is a 12-layer HubiTTSModel (d_model=512, 16 audio codebooks, vocab 1025) wired to maanka2/MOSS-Audio-Tokenizer-Nano

Early traction is already non-zero for a day-old card: HF reported ~772 downloads by this scout while likes were still at zero.

Run It

pip install torch torchaudio transformers huggingface_hub

import torch
from transformers import AutoModel

device = "cuda" if torch.cuda.is_available() else "cpu"
model = AutoModel.from_pretrained(
    "ayeeyokoris/hubi-tts",
    trust_remote_code=True,
).to(device)

model.tts("Salaam, sidee tahay!", save_path="output.wav")
cloned = model.tts(
    "Speaking with a cloned voice.",
    prompt_audio="reference.wav",
    save_path="cloned.wav",
)

The bundled modeling_hubi_tts.py / processing_hubi_tts.py register through auto_map, so the same AutoModel / AutoProcessor path the README shows is the intended day-0 surface.

Notes for makers

  • Expect a custom-code install (trust_remote_code=True) and a pull of the MOSS audio tokenizer dependency named in config
  • Best first test is short Somali lines from the README before you lean on cloning quality
  • No Gen CTA: Somali open TTS does not map to the Flux / Grok / MiniMax-H3 / Seedream / Seedance Gen loop

Conclusion

If you need a small, clone-capable Somali TTS checkpoint you can run from HF today, Hubi-TTS is the zero-day card to try — then decide if the missing SPDX line and custom-code trust are acceptable for your stack.

—Anabel ♡