Cairo-TTS · كايرو

Fast Egyptian Arabic text-to-speech with English code-switching, streaming and zero-shot voice cloning.
It runs on CPU or GPU, starts speaking in under 100 ms on a GPU, and is about 100M parameters.

---

Highlights

🗣️ Egyptian Arabic	Trained on about 747 hours of Egyptian podcast speech. It reads everyday عامية naturally.
🔀 Code-switching	English words written in Latin letters inside Arabic sentences: «ال meeting بكره الصبح».
🎙️ Zero-shot cloning	Any voice from a 5–10 s recording, plus 20 preset voices (10 male, 10 female).
⚡ Streaming	Audio is decoded in 80 ms frames while it is generated.
💻 CPU or GPU	Faster than real time on a laptop CPU.
🔢 Text frontend	Numbers, dates, times, money, percentages, phone numbers, emails, URLs and acronyms are read out in Egyptian.
✍️ Pronunciation control	Optional diacritics override the model's own reading.

Quick start

pip install cairo-tts          # add [ui] for the web interface

from cairo_tts import CairoTTS

tts = CairoTTS.from_pretrained("mohammedaly22/Cairo-TTS")          # CPU by default; device="cuda" for a GPU
tts.save("ال meeting بكره الصبح بدري, ف يا ريت متتأخرش.", "out.wav", voice="malak")
tts.save("صوتي انا بقى.", "me.wav", voice="my_recording.wav")      # zero-shot cloning
for chunk in tts.stream("بث مباشر وانت بتسمع."):                    # float32 chunks @ 24 kHz
    ...
print(tts.last_metrics)                                             # TTFA, total time, real-time factor

cairo-tts say "النهارده الجو حلو اوي." -o out.wav --voice ahmed
cairo-tts serve                     # web UI on http://127.0.0.1:8000

Pronunciation control

The model diacritizes on its own. Diacritics you write override it:

You write	Effect	Example
a vowel	picks between readings	نِفسَها (she wants) vs نَفسَها (herself)
ّ / ْ	gemination / no vowel	بَتمَشَّى, اتْحَرَجْت
Latin words	read in English	ال deadline, ال project
CEO, AI	upper-case runs are spelled as English letters	→ c e o, a i

Diacritics may be written in any order; the frontend normalizes them. . ends a sentence (longer pause), , or ، makes a short pause, and ? / ! become ..

Preset voices

Twenty preset voices are in voices/, described in voices/voices.json:

• male: محمد، أحمد، علي، عمر، يوسف، مصطفى، كريم، حسن، طارق، يحيى
• female: سارة، مريم، نور، فاطمة، هبة، ياسمين، منى، رنا، دينا، ملك

Each is a clean 5–10 s reference. Fourteen (mohamed, ahmed, ali, youssef, mostafa, karim, hassan, tarek, yehia, sara, mariam, yasmin, mona, rana) are recordings contributed by the author; six (omar, nour, fatma, heba, dina, malak) are speakers from the training podcasts, chosen by cloning fidelity and verified against each speaker's voice print. All were mastered to the same profile: 70 Hz high-pass, background noise gated out of the pauses, −20 LUFS, peaks ≤ −3 dBFS. tarek is the default voice. The names are made up.

Model details

Architecture	Pocket TTS (Kyutai): a causal transformer over Mimi latents with a 1-step flow (LSD) head, then a streaming Mimi decoder at 24 kHz, 12.5 frames/s
Size	6-layer backbone, about 100M parameters, 382 MB (fp32 safetensors)
Tokens	Arabic letters, diacritics, a–z, , .
Training data	About 747 hours of Egyptian podcast speech: diarized, quality-filtered, transcribed and loudness-normalized to −20 LUFS
Recipe	24-layer teacher fine-tuned from Pocket TTS (64k steps, cosine cooldown), distilled into this 6-layer student (30k steps); voice prompts ≤ 5 s
Code	Inference: Cairo-TTS

Audio quality

The 1-step flow sampler loses some high-frequency detail (median −2 dB at 1–2.5 kHz, −5 to −7 dB at 3–6 kHz versus the cloned reference, measured on 20 voices). The cairo-tts package corrects it with a linear-phase filter (clarity=1, the default): DNSMOS OVRL 3.336 → 3.359, SIG 3.548 → 3.570, bandwidth 5.9 → 8.0 kHz, with noise (BAK) and UTMOS unchanged. clarity=0 gives the raw output.

Speed

Measured with the cairo-tts package, after warm-up, on 7–11 s sentences:

Device	Time to first audio	Speed
NVIDIA GTX 1660 Ti (6 GB)	about 80 ms	about 2.2× real time
Intel i7-10750H (laptop CPU)	about 330 ms	about 2.6× real time

Faster GPUs mostly shorten the time to first audio. The model is small, so a laptop CPU is a good target.

Generation parameters

Parameter	Default	Meaning
temperature	0.3	sampling temperature: lower is steadier, higher is more expressive
lsd_steps	1	flow decoding steps per 80 ms frame
eos_threshold	−4.0	end-of-speech threshold: higher stops earlier
noise_clamp	off	clamps the sampling noise
frames_after_eos	auto	80 ms frames kept after the end of speech
max_letters	160	long text is split at sentence and comma boundaries into chunks of this many letters
clarity	1.0	correction of the model's measured high-frequency loss (0 = raw output)

Files

model.safetensors        weights
config.yaml              pocket-tts model config (relative paths)
tokenizer.json           79-symbol character vocabulary
ta_marbuta_lexicon.json  words whose final ه is a ta marbuta (ة)
voice.wav / .safetensors fallback voice (the default preset is voices/tarek.wav)
voices/                  20 preset voices + voices.json

Limitations

• Egyptian Arabic first. MSA works but sounds Egyptian, and other dialects are not supported.
• English words are pronounced the way Egyptian podcasters say them. Long English-only passages are weaker.
• Rare names and words can be misread. Add diacritics to fix them.
• Cloning copies the microphone and the room as well as the voice. Use clean, single-speaker prompts.
• Text longer than about 160 letters is generated chunk by chunk, so prosody does not carry across chunks.

Responsible use

This model can imitate voices. Do not use it to impersonate anyone without their consent, to deceive, to commit fraud, or to make content that appears to come from a real person. The preset voices come from public podcasts and are provided for research and demonstration only. Disclose synthetic audio when you publish it.

License

The code is MIT. The weights are fine-tuned from Kyutai's Pocket TTS and follow the Pocket TTS terms.

Citation

@misc{cairotts2026,
  title  = {Cairo-TTS: Fast Egyptian Arabic Text-to-Speech with Code-Switching and Voice Cloning},
  author = {Mohammed Aly},
  year   = {2026},
  url    = {https://huggingface.co/mohammedaly22/Cairo-TTS}
}

Thanks to Kyutai for Pocket TTS and Mimi.