# Run Lahgtna Egyptian-Arabic TTS on Google Colab

Generate Egyptian-Arabic speech with **v3** (multi-voice, non-diacritized) on a **free Colab GPU**.
Copy each cell into a Colab notebook (set **Runtime → Change runtime type → T4 GPU** first).

> The model repos are **private**, so you'll log in with your Hugging Face token (a **read** token
> from <https://huggingface.co/settings/tokens>).

---

### 1) Install
```python
!pip install -q omnivoice catt-tashkeel num2words soundfile
```

### 2) Log in to Hugging Face (needed for the private repo)
```python
from huggingface_hub import login
login()   # paste a READ token when prompted
```

### 3) Load the model + the bundled reference voice
```python
import torch
from omnivoice.models.omnivoice import OmniVoice
from huggingface_hub import hf_hub_download

REPO = "ehabnegm/lahgtna-omnivoice-egyptian-v3"
model = OmniVoice.from_pretrained(REPO, device_map="cuda", dtype=torch.float16)

ref_audio = hf_hub_download(REPO, "reference.wav")        # narrator voice clip
ref_text  = "كان العمل التطوعي واللي لما تفتح الباب بس ليه الناس"
```

### 4) Text front-end + long-form helper (chunk → no drift)
```python
import re, numpy as np, soundfile as sf
from num2words import num2words

_DIGITS = str.maketrans("٠١٢٣٤٥٦٧٨٩", "0123456789")
HARAKAT = re.compile(r"[ً-ْٰـ]")

def prep(t):                                   # numbers -> words, strip diacritics/English
    t = t.translate(_DIGITS)
    t = re.sub(r"\d+", lambda m: " " + num2words(int(m.group()), lang="ar") + " ", t)
    t = re.sub(r"[A-Za-z]+", " ", t)
    t = HARAKAT.sub("", t).replace("ى", "ي")
    return re.sub(r"\s+", " ", t).strip()

def chunks(t):                                 # split into clauses
    out, cur = [], ""
    for p in re.split(r"([،؛\.؟!\n]+)", t):
        cur += p
        if re.search(r"[،؛\.؟!\n]", p):
            if cur.strip(): out.append(cur.strip())
            cur = ""
    if cur.strip(): out.append(cur.strip())
    return out or [t]

def synth(text, num_step=16):                  # ALWAYS chunk long text
    parts = [model.generate(text=prep(c), language="arz",
                            ref_audio=ref_audio, ref_text=ref_text, num_step=num_step)[0]
             for c in chunks(text) if prep(c)]
    return np.concatenate(parts)
```

### 5) Generate & play
```python
from IPython.display import Audio
text = "أهلاً بحضرتك، معاك المساعد الذكي. اتفضل قوللي محتاج إيه، وأنا تحت أمرك."
audio = synth(text)
sf.write("out.wav", audio, 24000)
Audio("out.wav")
```

### 6) Long paragraph (works thanks to chunking)
```python
long_text = ("في السنين الأخيرة، بقى الذكاء الاصطناعي جزء أساسي من حياتنا. "
             "بس السؤال المهم، هو إحنا فاهمين التكنولوجيا دي كويس، ولا بس بنستخدمها وخلاص؟ "
             "خليني أوضّحلك حاجة مهمة، الذكاء الاصطناعي مش سحر، هو نماذج بتتعلم من البيانات.")
sf.write("long.wav", synth(long_text), 24000)
Audio("long.wav")
```

---

## Tips
- **Speed:** lower `num_step` → faster first audio (8 ≈ 0.5 s to first chunk, 4 ≈ 0.3 s) with a small quality trade-off.
- **Voice:** swap `reference.wav` for any 5–10 s clean clip of the voice you want to clone (and set `ref_text` to its transcript).
- **Language id:** always `language="arz"` (Egyptian Arabic).
- **Rule:** never send more than ~one sentence to a single `generate()` call — always chunk (the `synth()` helper does it for you).
- Other models: `ehabnegm/lahgtna-omnivoice-egyptian` (v1), `-v2` (v2). For v1/v2 keep diacritics
  (use the `catt-tashkeel` diacritizer) instead of the `HARAKAT.sub` stripping above.