ComfyUI-Antalia

Antalia-2 Mini Turkish TTS for ComfyUI

Turkish text to speech inside ComfyUI with Antalia-2 Mini (cloud0day3/antalia-mini, Apache-2.0): a small model (7.6M parameters, 46 MB for both weight files) with one synthetic male voice, 48 kHz output, and synthesis far faster than real time, on the CPU as well. It reads raw Turkish text: numbers, dates, times, amounts, initialisms, web addresses and English or brand words are turned into their spoken form by the package's own normaliser. The nodes call the official antalia-mini package and never download anything.

Unofficial. This node pack is an independent integration. It is not affiliated with or endorsed by PatientDesk AI (the authors of Antalia-2 Mini) or by Comfy Org.

日本語の概要は README.ja.md にあります。

What it does

| Node | Purpose | | --- | --- | | Antalia-2 Mini Loader | Loads the model from models/antalia/<name>/ (fp32 or fp16 weights) onto ComfyUI's GPU or the CPU | | Antalia-2 Mini Speak (Turkish TTS) | Text → audio (mono AUDIO), spoken_text (the text as the model reads it), report |

• seed: the same text, seed and settings give the same audio (on the same device).
• speed 0.95, steps 8, cfg 2.0 are the release defaults; 4 steps are faster and measurably less accurate.
• sample_rate: 48000 (native), or 24000 / 16000 / 8000 (telephone), resampled by the package.
• eq: the release's output EQ and de-esser (on by default); off gives the raw vocoder output.
• Long texts are split into sentence-sized chunks by the package and joined with short pauses.
• spoken_text shows the normaliser's reading, e.g. Saat 15:30'da → saat on beş otuzda.

Connect Save Audio (Advanced) or Preview Audio after it, or use the audio in a video workflow.

| Measured here (default settings, model loaded) | 3.4 s sentence | 17.7 s paragraph | | --- | --- | --- | | CPU (24 threads) | 0.06 s | 0.26 s | | RTX 5090 | 0.06 s | 0.14 s |

The authors' evaluation (model card): intelligibility level with the other small Turkish model EMA Lightning (WER 1.02 vs 1.00 % on Freya-TR-Eval), better in the 8 kHz telephone band, slightly lower full-band UTMOS.

How far it was checked

Windows 11, ComfyUI v0.42.1, CPU and RTX 5090; details and raw numbers in docs/VERIFICATION.md:

• Bit-identical to the official package. Seven cases (dates and times, a paragraph with an amount, a web address and English words, 4 and 1 steps, no guidance, 16 and 8 kHz, EQ off, fp16 weights): the pack's audio equals that of Antalia(model=<folder>).say(...) run in a separate environment, sample for sample, 7 of 7 on the CPU and 7 of 7 on CUDA. The official package itself gives identical samples run to run on both.
• Real ComfyUI queue (CPU and RTX 5090): the shipped workflows through Save Audio (FLAC) give the official audio to within half a 16-bit step (the FLAC rounding), at 48 kHz and 8 kHz, with fp16 weights and another seed; the spoken text is the official normaliser's.
• Clean install and CI (GitHub Actions; Ubuntu and Windows; ComfyUI v0.38.0 and v0.42.1): the real model on the CPU, compared with the official package, ComfyUI's validation of the workflows, and a real ComfyUI queue.

Not checked: speech quality itself (the pack reproduces the official output; see the model card's evaluation), Apple MPS, Linux and macOS outside CI.

Install

1. ComfyUI with the V3 node API (v0.38.0 or later) and Python 3.10+. A GPU is optional.

2. Clone this repository into ComfyUI/custom_nodes/ and install its requirement (antalia-mini==1.0.1, which brings cmudict and soundfile) into ComfyUI's Python:

   git clone https://github.com/hiroki-abe-58/ComfyUI-Antalia custom_nodes/ComfyUI-Antalia
   python -m pip install -r custom_nodes/ComfyUI-Antalia/requirements.txt
   

3. Download the model (46 MB; --variant fp32 or fp16 for one of them). The script pins the revision and verifies every file's SHA-256:

   python custom_nodes/ComfyUI-Antalia/tools/download_model.py --models-dir models/antalia
   

   Result: models/antalia/antalia-mini/ with config.json, acoustic(.fp16).safetensors, vocoder(.fp16).safetensors and the model card, LICENSE and NOTICE. Another location works through extra_model_paths.yaml with the key antalia.

4. Restart ComfyUI and open workflows/gui/speak.json (48 kHz FLAC and the spoken text) or speak_phone.json (8 kHz).

Notes and limits

• One voice, Turkish. English words inside Turkish text are read the way a Turkish speaker says them; whole English sentences are not what the model is for.
• The voice is synthetic (the model card: trained on speech generated with VoxCPM2, no recordings of real speakers). Please tell listeners when they hear a synthetic voice, as the authors ask.
• CPU and GPU give slightly different audio (43-64 dB SNR between them in the seven cases); each repeats itself exactly.
• Empty text, or text without anything readable, is an error rather than empty audio.

Licenses

• This node pack's code: MIT (LICENSE).
• antalia-mini (pip dependency) and the model (downloaded only by tools/download_model.py): Apache-2.0, PatientDesk AI (NOTICE).

Development

python -m pip install pytest ruff
python tools/download_model.py --models-dir <dir>/antalia
ANTALIA_MODELS_DIR=<dir>/antalia ANTALIA_COMFYUI_DIR=<ComfyUI checkout> python -m pytest tests
python scripts/ci_queue_check.py --comfyui <ComfyUI checkout> --models-dir <dir>/antalia [--gpu]   # real queue