Introduction

Matoub-4B just landed as a runnable Kabyle speech foundation. On 2026-11-05 ~21:23 UTC, agbalu/Matoub-4B published (GGUF companion minutes later), bringing zero-shot reference-conditioned voice cloning to Taqbaylit (kab, Latin script) at broadcast-style 24 kHz.

This is generative text-to-speech for a low-resource Amazigh language — not another English ASR pack. Give Kabyle text plus a short reference clip with its exact transcript, and the model aims to keep the speaker’s timbre while speaking fluent Taqbaylit.

What Shipped

  • Vendor / product: agbalu — Matoub-4B (Higgs-TTS 3 lineage)
  • Modality: open-weight Kabyle text-to-speech with native and cross-lingual voice cloning
  • Open vs closed: weights under the Boson Higgs TTS 3 Research and Non-Commercial License (research / personal / creator-content grant with attribution; commercial product use needs a separate Boson seat — not Apache/MIT)
  • Runnable today: transformers AutoModelForTextToWaveform with trust_remote_code=True on CUDA; NF4 4-bit path called out for ~8 GB laptop GPUs
  • Concrete fact: trained on 384,960 curated Common Voice Kabyle pairs; validation reports 0.881 mean WeSpeaker voice cosine across 16 utterances / six speakers, with CER excess 0.045 over the real-audio floor (Fadhma-300M)

The card frames Matoub as adapted from bosonai/higgs-tts-3-4b with an 8-codebook 24 kHz codec — spirantised, pharyngealised, emphatic, and geminated Kabyle consonants included.

Run It

pip install "transformers==5.17.0" "torch==2.11.0" "torchaudio==2.11.0" soundfile onnxruntime

import soundfile as sf
from transformers import AutoModelForTextToWaveform

REPO = "agbalu/Matoub-4B"
model = AutoModelForTextToWaveform.from_pretrained(
    REPO, trust_remote_code=True
).to("cuda").eval()

speech = model.generate(
    text="Azul fell-awen! Ass-a ad nemmeslay s teqbaylit.",
    reference_audio="voice.wav",
    reference_text="The exact words spoken in voice.wav, in their original language.",
)

sf.write("speech.wav", speech.waveform[0].numpy(), speech.sampling_rate, subtype="FLOAT")

For tight VRAM, the README documents NF4 loading so the transformer backbone compresses while the vocoder stays high precision. Expect a multi-GB download (model.safetensors ≈ 8.75 GB on the card tree).

Notes for makers

  • Reference audio can be English, French, Arabic, or another supported language — keep reference_text in the original language of that clip, not a Kabyle translation
  • Prefer clean single-speaker references without music or heavy reverb
  • Evaluation helpers on-card: Fadhma-300M CER, WeSpeaker cosine, TorchAudio SQUIM — useful for rejecting bad candidates, not a full naturalness arena
  • License is non-commercial / research by default with a creator-content attribution path; do not assume product redistribution freedom
  • No Gen CTA: Kabyle voice cloning does not map to the Flux / Grok / MiniMax-H3 / Seedream / Seedance Gen loop

Conclusion

If you need a paste-ready Kabyle voice-cloning stack today — and you can live with Boson’s research/NC terms or a commercial seat — Matoub-4B is the unlock: 4B, 384,960 CV pairs, 0.881 mean speaker cosine, transformers generate at 24 kHz.

—Anabel ♡