How to Run Antalia-2 Mini Locally: a 7.6M-Parameter Turkish TTS Model

Antalia-2 Mini is an open-weight Turkish text-to-speech model from cloud0day3. It has 7.62 million parameters, outputs 48 kHz audio, and runs faster than real time on a plain CPU, no GPU needed. It reads raw Turkish text, including numbers, dates and abbreviations, and speaks it in a single synthetic voice.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts and hardware requirements against the primary source before recommending an install path.

A new top-20 entrant on Hugging Face's trending list

Antalia-2 Mini entered the top 20 of Hugging Face's trending text-to-speech list. At time of writing it has 36 likes and 1,369 downloads on the cloud0day3/antalia-mini model page, released October 8, 2026, under the Apache 2.0 license. The top linked community Space is tardellirs/model-pulse (199 likes).

Specs at a glance

What it actually does

Antalia-2 Mini takes raw Turkish text and reads it in a single synthetic male voice, not a recording of a real person. A text normalizer built into the package converts numbers, dates, times, money amounts, abbreviations and English or brand names into the way a Turkish speaker would say them out loud, so you can hand it ordinary text instead of pre-cleaned, phonetic input. The model reads characters directly rather than phonemes, leaning on how consistently Turkish spelling maps to sound.

Speed: CPU first

Antalia-2 Mini is built to run comfortably without a GPU. Its real-time factor, the fraction of a clip's own length it takes to generate, is 0.052 on an 8-core CPU, meaning a 10-second clip renders in about half a second. On an Apple M5 it drops to 0.019 with 0.06-second first-audio latency when streaming, and on an Nvidia L40S GPU it reaches 0.026. The gap between CPU and GPU is small enough that most deployments can skip the GPU entirely.

Install it

The Python package covers normalization, synthesis, streaming and batched calls. It needs Python 3.10 or newer, and pulls in PyTorch, NumPy, safetensors, soundfile and cmudict as dependencies.

from antalia_mini import Antalia

tts = Antalia()  # downloads the weights (~31 MB) once; uses a CUDA GPU if there is one, else the CPU
speech = tts.say("Randevunuz 14 Ekim Salı günü saat 15:30'da.", path="randevu.wav", seed=0)
print(speech.duration, speech.sample_rate, speech.seed)

Passing a seed makes the same text and settings return the same audio every time; leave it out and the model picks one at random, returned in speech.seed. For long text, tts.stream(text) synthesizes a few sentences at a time instead of waiting on the whole thing, and Antalia(sample_rate=8000) resamples down to telephone or ASR-pipeline rates of 24000, 16000 or 8000 Hz. On Apple Silicon, Antalia(device="mps") runs on the GPU instead of the default CPU.

Known limitations

The model card is specific about where this falls short. It supports Turkish only, with one voice and no emotion control beyond speed. Short interjections come out weak, and text starting with the letters "ö" or "ı" can pick up a stray "l" sound at the very start, something the package works around automatically by default. Prosody does not carry across sentence boundaries, and unusual abbreviations, rare foreign names and Roman numerals may be read incorrectly. The model card also asks that anyone deploying it tell listeners they are hearing a synthetic voice, and rules out impersonation, fraud or other deceptive use.

FAQ

What is Antalia-2 Mini?

Antalia-2 Mini is an open-weight Turkish text-to-speech model from cloud0day3 with 7.62 million parameters. It converts raw Turkish text, including numbers, dates and abbreviations, into 48 kHz speech in a single synthetic voice.

Do I need a GPU to run it?

No. It runs well faster than real time on an 8-core CPU, with a real-time factor of 0.052. A GPU helps slightly but is not required.

Does it support languages other than Turkish?

No, it is Turkish only, with one fixed voice and no language switching.

How accurate is it?

On the Freya-TR-Eval public benchmark it scored a 1.09% word error rate and a 0.220% character error rate, the fewest character errors of the 15 Turkish TTS systems tested and the second lowest word error rate.

What license is it under?

Apache 2.0, covering the code, the weights and the text normalizer.

Where to go from here

• Need an English-first alternative instead? Run Kokoro-82M locally covers another small, fast open-weight TTS model built for broader language coverage.
• Want voice cloning instead of a fixed voice? Run Breeze TTS 2 locally covers a model built for voice clone, voice design and voice direction.
• Browse other speech and audio models: the local AI model directory lets you compare specs side by side.