𓋹 VoiceTut-TTS

The best open-source text-to-speech model for Egyptian Arabic & code-switching

🎧 Listen: hear VoiceTut-TTS vs. the base OmniVoice on Egyptian Arabic & code-switching → Audio demos · 🚀 Try it live: HuggingFace Space

VoiceTut-TTS is an Egyptian-Arabic text-to-speech system fine-tuned from OmniVoice on ~380 hours of Egyptian podcast speech. It produces natural Egyptian speech with seamless Arabic ↔ English code-switching, ships 15 built-in studio voices, supports zero-shot voice cloning, and includes a robust Egyptian-Arabic text normalization pipeline plus true streaming for long text.

Why "VoiceTut"? Tut — after the boy-king Tutankhamun (توت عنخ آمون) — anchors the model in Egyptian identity, just as our companion ASR model QwenCleo-ASR is named after Cleopatra. Together they form an Egyptian speech stack: Cleo listens, Tut speaks. 🎙️🗣️

✨ Features

• 🎯 Egyptian-first — fine-tuned specifically on Egyptian Arabic, not generic MSA.
• 🔀 Code-switching — handles real Arabic + English mixed speech (عندي meeting بكرة).
• 🗣️ 15 built-in voices — male & female studio speakers, each with style tags.
• 🧬 Zero-shot cloning — clone any voice from a few seconds of reference audio.
• 🔢 Robust normalization — numbers, dates, times, currencies, phones, emails, URLs, abbreviations + a diacritics override table and a custom lexicon.
• ⚡ True streaming — long text is split into sentences and yielded as audio chunks for low time-to-first-audio.
• 📦 pip-installable — pip install voicetut-tts (+ OmniVoice from GitHub), or clone and run locally.

📊 Performance

Measured on a single NVIDIA T4 (Colab), float16, num_step=32. Reproduce with examples/04_evaluation.ipynb.

Measured on a T4; expect markedly lower RTF / TTFA on an A100 / H100. RTF and TTFA scale with num_step — drop to num_step=16 for faster, slightly lower-quality output.

🗣️ Sample Outputs

🎧 Listen to all built-in voices & VoiceTut vs. base OmniVoice comparisons → — pure Egyptian, code-switching, English, normalization, per-speaker, customer-service, and zero-shot cloning samples. Or try them live in the HuggingFace Space.

📦 Installation

Option A — PyPI

# 1. PyTorch matching your CUDA (see https://pytorch.org)
pip install torch --index-url https://download.pytorch.org/whl/cu121
# 2. The OmniVoice backbone (not on PyPI, so install it from GitHub)
pip install git+https://github.com/k2-fsa/OmniVoice.git
# 3. VoiceTut-TTS
pip install voicetut-tts
# optional: web UI deps
pip install "voicetut-tts[web]"

Option B — from source

git clone https://github.com/MohammedAly22/VoiceTuT-TTS.git
cd VoiceTuT-TTS
conda create -n voicetut python=3.10 -y && conda activate voicetut
pip install torch --index-url https://download.pytorch.org/whl/cu121
pip install git+https://github.com/k2-fsa/OmniVoice.git    # backbone
pip install -e ".[web,dev]"

🚀 Usage

Python API

from voicetut_tts import VoiceTutTTS

tts = VoiceTutTTS.from_pretrained("mohammedaly22/VoiceTut-TTS")

# 1) Built-in speaker
tts.synthesize("ازيك عامل ايه النهاردة؟", speaker="Mohamed", output="out.wav")

# 2) Zero-shot voice cloning
tts.synthesize("النهارده الجو حلو اوي",
               ref_audio="my_voice.wav", ref_text="ده الصوت بتاعي",
               output="clone.wav")

# 3) Generation parameters
tts.synthesize("عندي meeting الساعة 3:30",
               speaker="Asmaa", num_step=48, guidance_scale=2.5, speed=1.1,
               output="cs.wav")

True streaming (long text)

import sounddevice as sd
for sr, chunk in tts.stream(long_paragraph, speaker="Sayed"):
    sd.play(chunk, sr); sd.wait()          # play each sentence as it's ready

# or write a single concatenated file
tts.synthesize_long(long_paragraph, "long.wav", speaker="Sayed")

Text normalization & custom lexicon

from voicetut_tts import ArabicNormalizer

norm = ArabicNormalizer()
norm("عندي 250 جنيه والميعاد 3:30 يوم 14/3/2024")
# -> "عندي ميتين وخمسين جنيه والميعاد تلاتة و نص يوم أربعتاشر مارس ألفين وأربعة وعشرين"

What gets normalized

Override tables & runtime dictionaries

# diacritized-form overrides (win over the CSV table)
tts.add_lexicon({"تيوت": "تُوت", "نايل": "نَايِل"})

# English-name -> Arabic for correct pronunciation
tts.add_names({"Ziad": "زياد", "Kareem": "كريم"})

Editable data tables ship with the package:

• data/diacritics.csv — word,diacritized (Arabic word → diacritized form)
• data/names_en_ar.csv — english,arabic (name transliteration)

CLI

voicetut --list-speakers
voicetut --text "ازيك عامل ايه؟" --speaker Mohamed --output out.wav
voicetut --text "نص طويل..." --speaker Sayed --stream --output long.wav

🌐 Serving

A custom-styled Gradio web UI (black/white + blue theme, speaker dropdown with gender + style tags, reference preview, voice cloning with mic recording, AR/EN language switch, generation params, examples):

pip install "voicetut-tts[web]"
python app.py                                   # default HF checkpoint, port 7860
OMNICLEO_CKPT=exp/omnivoice_egy/checkpoint-8000 python app.py   # local checkpoint
OMNICLEO_SHARE=1 python app.py                  # public share link

📓 Examples (Colab)

Notebook	Description
01_quickstart.ipynb	Install, load, synthesize with a built-in voice
02_voice_cloning.ipynb	Zero-shot cloning + normalization & lexicon
03_web_ui.ipynb	Launch the Gradio web UI from Colab (public link)
04_evaluation.ipynb	Measure RTF, TTFA, VRAM, WER, similarity, UTMOS

🔗 Links

• 🤗 Model: https://huggingface.co/mohammedaly22/VoiceTut-TTS
• 📦 PyPI: https://pypi.org/project/voicetut-tts/
• 🧠 Base model: https://github.com/k2-fsa/OmniVoice
• 🎧 Companion ASR: https://github.com/MohammedAly22/qwencleo-asr

📜 License & Citation

Released under the Apache-2.0 license.

@software{voicetut_tts_2026,
  author  = {Mohammed Aly},
  title   = {VoiceTut-TTS: Egyptian Arabic & Code-Switching Text-to-Speech},
  year    = {2026},
  url     = {https://github.com/MohammedAly22/VoiceTuT-TTS},
  note    = {Fine-tuned from OmniVoice}
}