Downloads · 30 days
0
ehabnegm/masri-higgs-v3-egyptian-lora
masri-higgs-v3-egyptian-lora is a text-to-speech model from ehabnegm. Use it when you need text read aloud. It is set up for peft. The card lists the license as other.
Everything needed to reproduce, audit, or extend the Egyptian Arabic TTS fine-tune of ehabnegm/higgs-tts-3-4b-egyptian-v3 on ehabnegm/masri-100h-egyptian-tts-enhanced.
Downloads · 30 days
0
Access
Public
Updated Aug 10, 2026
Repo size
1.1 GB
Likes
0
Public
Click a slice to open those files.
.safetensors1.1 GB · 98%
From the Hugging Face model README
Everything needed to reproduce, audit, or extend the Egyptian Arabic TTS fine-tune of
ehabnegm/higgs-tts-3-4b-egyptian-v3
on ehabnegm/masri-100h-egyptian-tts-enhanced.
Want the ready-to-serve model instead? → ehabnegm/masri-higgs-v3-egyptian-tts
adapters/step600/ ← recommended: best generalization (505 MB)
adapters/step1134/ ← final step: lowest val loss, most voice-adapted (505 MB)
scripts/ ← encode, train, evaluate, re-score, build listening pages
eval/ ← all raw evaluation JSON + validation-loss curve
samples/ ← 45 WAVs: base vs step600 vs step1134 on out-of-domain text
| step600 | step1134 | |
|---|---|---|
| Val loss | 4.0995 | 4.0875 |
| CER median | 0.0145 | 0.0198 |
| WER mean | 0.1012 | 0.1219 |
| Best for | production / clarity | maximum voice adaptation |
Base is CER 0.0300 / WER 0.1146 / val 4.2168. Note step1134's WER is worse than base even though its loss is lowest — validation loss kept improving while round-trip intelligibility degraded after step 600. That's mild overfitting to the training distribution, and it's why step600 ships.
import sys, torch
sys.path.insert(0, "serving") # from the base model repo
from higgs_v3_model import HiggsV3ForTTS
from peft import PeftModel
model, _ = HiggsV3ForTTS.from_checkpoint("path/to/higgs-tts-3-4b-egyptian-v3", dtype=torch.bfloat16)
model = PeftModel.from_pretrained(model, "adapters/step600")
model = model.merge_and_unload().cuda().eval()
# 1. audio -> 8-codebook codes @ 25 Hz (98 h in ~14 min on an A100, 410x realtime)
python3 scripts/encode_codes.py
# 2. LoRA fine-tune (3 epochs, 1134 steps, ~4 h on an A100-80GB)
python3 scripts/train_lora.py --epochs 3 --bs 32 --accum 2 --rank 64 --lr 1e-4
# 3. evaluate: synthesize held-out text, transcribe back, score
python3 scripts/eval_roundtrip.py --adapter ckpt/step600 --n 40 --tag step600
# 4. out-of-domain production test across all checkpoints
python3 scripts/eval_multi.py --ckpts base,step200,step400,step600,step800,step1134
Training. LoRA r=64, α=128, dropout 0.05 on q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
— 132.1 M trainable of 4.176 B (3.16%). bf16 + gradient checkpointing, batch 32 × accum 2,
OneCycle LR 1e-4, 3% warmup. Peak VRAM 21.4 GB. Validation is video-disjoint (6 held-out source
videos), so no episode leaks between train and eval.
Loss is next-frame teacher forcing over the 8 codebooks with the delay pattern applied
(codebook c shifted down by c rows, BOC leading / EOC trailing), cross-entropy with
ignore_index=-100 masking text and reference positions.
Evaluation. There is no reference audio for unseen text, so quality is measured by
ASR round-trip: synthesize → transcribe with oddadmix/whisper-large-v3-turbo-arabic-dialectal-v2
→ CER/WER against the input text. All checkpoints use the same voice reference and a fixed seed per
sentence, so differences are model differences, not sampling luck.
| step | 0 | 200 | 400 | 600 | 800 | 1000 | 1134 |
|---|---|---|---|---|---|---|---|
| val loss | 4.2168 | 4.1379 | 4.1171 | 4.0995 | 4.0912 | 4.0878 | 4.0875 |
Monotonic, no divergence, clearly flattening — a 4th epoch would not have paid.
Derived from a real, identifiable person's voice from public YouTube content. Do not publish, commercialize, or impersonate without consent. Disclose synthetic audio.