Skip to main content
Back

Antalia-2 Mini Beats 2.38B Models on Turkish Speech at Just 30MB

A 7.6M-parameter Turkish text-to-speech model runs 50x faster than real time on a laptop CPU, with the fewest character errors among 15 Turkish systems benchmarked.

Model, Full Access
  • PatientDesk AI released Antalia-2 Mini, a 7.62M-parameter Turkish TTS model under Apache-2.0.
  • Lowest CER (0.19%) among 15 Turkish systems on Freya-TR-Eval, matching models 300x larger.
  • Runs 50x faster than real time on an Apple M5 CPU with 2 threads, no GPU needed.
  • Flow matching with shortcut self-consistency lets one model sample in 1, 2, 4 or 8 steps.
  • Trained on 1,578 hours of synthetic speech generated by VoxCPM2, no real speakers used.
  • Install with pip install antalia-mini; browser demo available on Hugging Face Spaces.

Antalia-2 Mini is an open-source Turkish text-to-speech model designed for local, CPU-only inference. Its 7.62 million parameters occupy 30.5 MB in fp32, and the project publishes the weights, text normalizer, and training recipe under Apache 2.0.

Turkish-language developers have often had to choose among hosted APIs, multi-gigabyte multilingual models, and compact systems with higher recognition errors. Antalia-2 Mini narrows its scope to one language and one voice, trading flexibility for a small footprint and low latency.

Antalia-2 Mini at a glance
Parameters 7.62 million
Weight size 30.5 MB in fp32
Language Turkish
Speaker support One fixed voice
Default sampling Eight refinement steps
Target runtime CPU, GPU, or browser
License Apache 2.0

A 7.62 million-parameter model near the top

Published results on Freya-TR-Eval place Antalia-2 Mini among the most accurate Turkish systems measured by word error rate and character error rate. WER and CER compare generated speech with the intended text after automatic transcription, with lower percentages indicating fewer recognition errors.

Reported Freya-TR-Eval results
Evaluation WER CER Placement
Default settings 0.97% 0.19% Tied with EMA Lightning on WER; lowest CER
Public benchmark rerun 1.05% 0.20% Second of 15 on WER; first on CER

The public rerun placed Antalia-2 Mini 0.01 percentage points behind EMA Lightning on WER, a difference the project describes as within seed-to-seed variation. The same leaderboard includes Trendyol-TTS at 2.38 billion parameters, Anka TTS at 336 million, Chatterbox Multilingual at 500 million, and Gemini 3.8 Flash TTS. Each recorded a higher CER than Antalia-2 Mini.

WER and CER measure intelligibility through a speech recognizer. Naturalness, voice preference, emotional range, and prosody require listening tests, while production performance can vary with names, domain-specific vocabulary, punctuation, and noisy source text.