Multi-voice Egyptian-Arabic (arz) TTS, fine-tuned on a combined ~34 h, two-narrator corpus. Speaks two trained voices (Eqkawkab + Noselleel) and clones any reference voice zero-shot. This repo is the hub for the project — it also hosts the benchmark, streaming interface, code, training guide, and sample sets.

For the best multi-voice pronunciation, see v3 (non-diacritized) — it beats v2 by ~38% lower CER on pure-Arabic.

👪 Model family

	repo	voices	text
base	lahgtna-omnivoice-v2	—	diacritized
v1	egyptian	1	diacritized
v2 ← you are here	egyptian-v2	2	diacritized
v3 (recommended)	egyptian-v3	2	raw
🎛️ Compare GUI	Space	—	—
📚 Data	eqkawkab · noselleel	—	—

What's in this repo

model.safetensors + config + tokenizer — the v2 model
benchmark/ — 4-way ASR-CER benchmark + 60 audio clips
STREAMING.md + samples_streaming/ — low-latency streaming (TTFC ~0.8s)
samples_v2_multivoice/, samples_v2_long/ — both-voice demos
TRAINING_GUIDE.md + code/ — full reproducible recipe + scripts

Usage

python

import torch, soundfile as sf
from omnivoice.models.omnivoice import OmniVoice
m = OmniVoice.from_pretrained("ehabnegm/lahgtna-omnivoice-egyptian-v2", device_map="cuda", dtype=torch.float16)
audio = m.generate(text="أهلاً بيك", language="arz", ref_audio="voice.wav", ref_text="نص المرجع")
sf.write("out.wav", audio[0], 24000)

Training

~34 h / 10,828 clips (Eqkawkab + Noselleel), diacritized text. Continued from v1 checkpoint-6000, 8,000 steps, SDPA, bf16, 1× L4. Train loss 3.93 → 1.92.

Benchmark (ASR round-trip CER, 13 pure-Arabic sentences — lower = clearer)

model	CER ↓
v1	0.056
v3	0.063
v2	0.102
base	0.109
Full benchmark + audio in `benchmark/`.

321-word long-form stress test

A full Egyptian-Arabic article (~321 words, ~2:50) synthesized by each model, then transcribed back with Whisper large-v3 (ASR round-trip CER — lower = stays on-text).

model	audio	RTF	CER ↓
v3	173 s	0.25	0.21 — stays coherent ✅
base	170 s	0.35	0.69 — drifts ⚠️
v1	170 s	0.35	0.71 — drifts ⚠️
v2	170 s	0.43	0.71 — drifts ⚠️

Finding: on paragraph-length input the diacritized models (base/v1/v2) lose coherence and babble; v3 (non-diacritized) is the only version that stays on-text — decisive for long-form and agent replies. Listen: base.wav, v1.wav, v2.wav, v3.wav.

Audio: long300/.

Limitations & responsible use

CATT diacritics on heavy colloquial are approximate; code-switched English is not trained. Audio derived from public YouTube channels (@Eqkawkab, @noselleel) — credit the creators, get permission before commercial use, and don't use the voices to impersonate or mislead.