VoiceTut-TTS

VoiceTut-TTS

Egyptian Arabic text-to-speech with Arabic and English code-switching, 19 built-in voices, voice cloning and streaming of long text. For developers and creators, using only voices they have the right to use.

Made by mohammedaly22 on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: VoiceTut-TTS is a text-to-speech model for Egyptian Arabic, fine-tuned from the OmniVoice model. It speaks natural Egyptian Arabic and switches between Arabic and English inside one sentence. It comes with 19 built-in voices, can clone a voice from a short reference clip, and streams long text sentence by sentence.

Who it suits: It suits developers and creators who need Egyptian Arabic speech, including mixed Arabic and English text such as workplace talk.

Industry | Who uses it | How they use it | Example
Media | Video and podcast creators | Voice narration in Egyptian Arabic | Narrating a short explainer video
Software | App developers | Add spoken replies to an app | Reading chat answers aloud in Egyptian Arabic
Customer service | Voice bot builders | Speak mixed Arabic and English phrases | Saying a meeting time that includes English words
Accessibility | Tool makers | Read long text aloud as it is written | Streaming a long article in chunks

You can skip it if you need other Arabic dialects, since the model was tuned for Egyptian Arabic rather than standard Arabic.

Where to use it: It runs as a Python library on Windows, macOS or Linux, with a GPU preferred. The card tested it on an NVIDIA T4 in Google Colab.

• Python scripts that import voicetut_tts
• A CUDA GPU, as in the card's tests
• A hosted demo Space on Hugging Face for a first listen

Use case | Who | What they get back
Built-in voice speech | Developers | A WAV file in one of 19 voices
Voice cloning | Creators | Speech in a voice taken from a short reference clip
Streaming long text | App builders | Audio yielded sentence by sentence

How to set it up: You need Python, PyTorch built for your CUDA version, the OmniVoice backbone, and the voicetut-tts package. The card says the OmniVoice backbone is not on PyPI, so you install it from its source repository. Use only voices you have the right to clone.

• Python
• PyTorch matching your CUDA version
• The OmniVoice backbone, installed from its source repository
• About 2.5 GB of disk for the F32 weights

1. Install PyTorch for your CUDA version.
2. Install the OmniVoice backbone from its source repository.
3. Install the voicetut-tts package with pip.

pip install voicetut-tts

How to use it: Create the model from the pretrained name, then call synthesize with text and either a built-in speaker or a reference clip. To stream long text, call stream, or use synthesize_long to write a single file.

from voicetut_tts import VoiceTutTTS

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

# 1) Built-in speaker
tts.synthesize("ازيك عامل ايه النهاردة؟", speaker="Mohammed", 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) Code-switching + generation params
tts.synthesize("عندي meeting الساعة 3:30 ومعايا ال presentation",
               speaker="Yasmin", num_step=48, guidance_scale=2.5, speed=1.05, output="cs.wav")

for sr, chunk in tts.stream(long_paragraph, speaker="Ahmed"):
    play(chunk)                    # plays each sentence as it's generated
tts.synthesize_long(long_paragraph, "long.wav", speaker="Ahmed")

Tool | What it does
synthesize | Writes speech to a file from text and a speaker or reference clip
stream | Yields audio chunks one sentence at a time
synthesize_long | Writes a long paragraph to a single file

Example: You call synthesize with an Egyptian Arabic greeting and the speaker Mohammed. It writes a WAV file in which that built-in voice says the greeting in natural Egyptian Arabic.

For newcomers and experts: If you are new, try the hosted demo first to hear the voices. Then install the package and run the first example to make one audio file.

If you know the field, the model adds Egyptian Arabic text normalization for numbers, dates, times, currencies, phone numbers, emails and abbreviations, plus an English to Arabic name map. On a T4 in float16 with 32 steps, the mean real-time factor is 1.13 and the best is 0.49.

Egyptian Arabic or mixed text
Text normalization
OmniVoice-based speech model
24 kHz audio, whole or streamed

Good to know: Quality numbers come from the card's own tests on one T4 GPU, with a time to first audio of 1.68 seconds when streaming.

• License: Apache 2.0
• Cost to run: no API key, account or paid service is named, but a GPU with about 3 GB of VRAM is used in the tests
• Data it reaches: the text and reference audio you provide, plus files downloaded from Hugging Face and GitHub
• Not for: cloning voices you do not have the right to use

Title: Word error rate, lower is better
Egyptian Arabic: 0.40
English: 0.07

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: omnivoice
• Parameters: 613M
• Weights: about 2.5 GB in F32; about 306 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 3,074
• Likes: 36
• License: apache-2.0
• Languages: ar, arz, en
• Based on: k2-fsa/OmniVoice
• Quantized versions: 1
• Fine-tunes: 2
• Tags: omnivoice, safetensors, text-to-speech, egyptian-arabic, arabic, code-switching, voice-cloning

Links

• Model: https://huggingface.co/mohammedaly22/VoiceTut-TTS
• Files: https://huggingface.co/mohammedaly22/VoiceTut-TTS/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download mohammedaly22/VoiceTut-TTS

Then follow the usage in its model card. Nothing in this kit runs by itself.