vadimbelsky/emirati-fastpitch-bilingual-v1.0

🇦🇪 Emirati FastPitch v1.0

A fine-tuned FastPitch text-to-speech model for Emirati Arabic with Arabic–English bilingual code-switching support.
Designed for natural prosody, dialectal Emirati pronunciation, and seamless AR/EN mixing in real-world speech.

🔊 Audio Samples

🇦🇪 Full Emirati Arabic

Reference text:
شو رأيك لو نعمل ورشة صغيرة للأطفال بالمدرسة عن الفنون اليدوية، ترى الأطفال يحبون الحركات هذي وبنشجعهم على الإبداع، إنت موافق تساعدنا بالمواد والأفكار ولا مب فاضي

---

🌍 Emirati + English (Code-Switching)

Reference text:
اليوم جلسنا وقسمنا المهام في الـ Project Management عشان نحافظ على المحميات، الـ AI يساعد في رصد الحيوانات والطيور والـ Dashboard يعطينا تنبيهات، شو تتوقعون بيصير لو ما سوينا Update سريع، ترى الوضع مهم وايد

---

🇺🇸 English Only

Reference text:
Basically, pair programming feels awkward at first, but you know, it actually speeds up problem solving and learning

💡 All samples were synthesized using Emirati FastPitch v1.0 + Emirati HiFi-GAN bilingual vocoder at 22,050 Hz.

✨ Features

• Emirati Dialect Focus: Trained on authentic Emirati Arabic speech
• Bilingual Phonemes: AR_ and EN_ prefixed phoneme inventory
• Code-Switching Ready: Natural transitions between Arabic and English
• Custom Tokenizer: Space-delimited, multi-character phoneme tokens
• High Prosody Control: Duration + pitch prediction (FastPitch)

---

🧠 Model Details

Item	Value
Model Type	FastPitch
Framework	NVIDIA NeMo
Sample Rate	22,050 Hz
Epochs	1,059
Training Data	10+ hours (Emirati Arabic + English)
Phoneme Inventory	200+ AR_/EN_ tokens

---

🔁 Vocoder (Required)

This model is designed to be used with:

Emirati HiFi-GAN Bilingual v1.0
https://huggingface.co/vadimbelsky/emirati-hifigan-bilingual-v1.0

The vocoder is trained on matching mel-spectrogram configuration, ensuring:

• clean consonants
• stable Arabic emphatics
• natural English sibilants
• smooth AR/EN transitions

---

🚀 Quick Start (NeMo)

1. Load FastPitch

from nemo.collections.tts.models import FastPitchModel

fastpitch = FastPitchModel.from_pretrained(
    "vadimbelsky/emirati-fastpitch-bilingual-v1.0"
)
fastpitch.eval()

2. Convert Text → Phonemes

from emirati_phoneme_converter import EmiratiPhonemeConverter

converter = EmiratiPhonemeConverter()
phonemes = converter.convert(
    "اليوم عندنا meeting مهم with the AI team"
)

3. Generate Mel-Spectrogram

tokens = fastpitch.parse(phonemes)
mel = fastpitch.generate_spectrogram(tokens=tokens)

4. Synthesize Audio with HiFi-GAN

from nemo.collections.tts.models import HifiGanModel

vocoder = HifiGanModel.from_pretrained(
    "vadimbelsky/emirati-hifigan-bilingual-v1.0"
)
vocoder.eval()

audio = vocoder.convert_spectrogram_to_audio(mel)

🌍 Supported Languages • Emirati Arabic (dialect) • English (US-style phonemes) • Mixed Arabic–English speech (code-switching)

📜 License

This model is released under CC-BY-4.0.

📌 Citation

Emirati FastPitch v1.0 – Bilingual Emirati Arabic TTS Vadim Belsky, 2026