Matoub-4B (GGUF)

Standalone GGUF weights for Matoub-4B, a reference-conditioned Kabyle speech synthesis model producing 24 kHz mono audio.

Provided quantizations

File	Size (decimal GB)	Use
`Matoub-4B-q4_k.gguf`	3.912	Recommended compact option on the measured development panel
`Matoub-4B-q8_0.gguf`	5.934	Higher-precision decoder option
`Matoub-4B-bf16.gguf`	9.724	Native runtime precision baseline

All formats preserve all 528 codec tensors in FP32, with values verified against the source. Quantization applies to eligible generator tensors. Required configuration and tokenizers are embedded. No separate codec download is needed.

Measurements

Measured on one H100 NVL MIG 3g.47gb, four text/reference cases and two seeds per case. Each request used one decoding attempt, a 600-position limit, temperature 0.8, top-p 0.8 and top-k 30.

Format	Peak device memory (GB)	RTF on completed outputs	Completed / attempted
BF16	9.904	0.290	5 / 8
Q8_0	6.694	0.284	4 / 8
Q4_K	4.716	0.138	5 / 8

RTF is synthesis time divided by generated audio duration; lower is faster. It includes reference processing and waveform decoding, excludes model loading and quality scoring, and is weighted by successful audio duration. Failed requests are counted separately. Different completion subsets prevent interpreting these values as a controlled speedup ratio. Memory is a sampled peak above idle device usage, not a minimum hardware requirement.

On the four requests completed by both Q4_K and native BF16 (280 normalized characters), Fadhma-300M CER was 7.14% versus 8.21%; WeSpeaker cosine was 0.542 versus 0.546. Predicted SQUIM quality was 4.180 versus 4.191. This small panel supports Q4_K as a compact option; it does not prove equivalent quality for all speakers or texts. Q8_0 is offered for precision preference, not as a measured quality winner. Q4_0 is excluded after weaker results.

Usage

Use audio.cpp, verified at v0.9.4, commit f7ac446fb75e594a6913c846bea6c85b637f6ac0, with CUDA on SM90. This speech GGUF requires that runtime's higgs_audio_tts family; a generic text-only GGUF loader is not sufficient. CPU and other accelerators were not qualified in this release.

audiocpp_cli --task clon --family higgs_audio_tts \
  --model Matoub-4B-q4_k.gguf --backend cuda \
  --text "Azul fell-awen! Ass-a ad nemmeslay s teqbaylit." \
  --voice-ref voice.wav \
  --reference-text "The exact words spoken in voice.wav." \
  --max-tokens 1500 \
  --session-option higgs_audio_tts.codec_weight_type=f32 \
  --out speech.wav

Use a clean single-speaker reference and its exact transcript. The native runtime has its own decoding path. It does not include the Transformers package's ASR, speaker verification or multi-candidate selection. The example's larger limit does not guarantee completion; generation may still repeat or reach the limit.

Main model and limitations

See Matoub-4B for the Python interface, intended use, pronunciation limitations and citation. Voice identity and pronunciation vary across reference recordings. Use voices with permission and identify synthetic speech appropriately.