Antalia-2 Mini: A 7.6-million-parameter Turkish text-to-speech model trained entirely on synthetic speech
Organizations: PatientDesk AI
Abstract
Antalia-2 Mini is a single-voice Turkish text-to-speech model with 7,620,931 parameters. It reads characters and produces 48 kHz audio at 41 times real time on an Apple M5 laptop CPU (2 threads). Its training audio came only from VoxCPM2 (the teacher), about 300 times larger, cloning one synthetic clip; automatic checks kept about 1,790 h. Fine-tuning the vocoder on the acoustic model's own mel spectrograms cut the word error rate (WER) of Whisper's transcripts from a mean of 3.89% to 1.75% at adjacent pre-release checkpoints. The teacher corrupts the start of a clip when the first word begins with ö, the student learned this, and a spoken prefix we cut from the output hides it. On the 495-sentence Freya-TR-Eval benchmark, where we tuned settings, version 1.0.1 has WER 1.09% and character error rate 0.22% on our public harness. A paired bootstrap cannot distinguish these from EMA Lightning, an 8.6M Turkish model, which scores higher on full-band UTMOS, an English-trained quality proxy. No human listening test exists. Weights, inference code and the benchmark harness are public under Apache-2.0; training code and data are not.
Figures & tables
| Teacher and reference | CER, % | UTMOS | Identity | F0, Hz | In band, % |
|---|---|---|---|---|---|
| VoxCPM2, chosen reference | 2.95 | 3.387 | 0.697 | 129.42 | 83.3 |
| VoxCPM2, earlier model’s render | 3.03 | 3.295 | 0.748 | 139.33 | 45.0 |
| VoxCPM2, second refinement | 2.82 | 3.365 | 0.631 | 125.13 | 50.0 |
| Alania-2 API, earlier model’s render | 3.12 | 2.648 | 0.623 | 146.71 | 23.3 |
| Rule | Rejects when | Relaxed in the short set | Long set | Short set |
|---|---|---|---|---|
| Length limits | Clip too long or too short | 8 | 13 | |
| CER | Above 10% | Up to 4 words: passes if the clip’s text appears in the transcript. Interjections: off | 14,227 | 289 |
| Speaking rate | Outside 8–24 characters per second | Up to 2 words: no minimum | 5 | 1,051 |
| ECAPA identity | Below 0.45 | Under 1.2 s: off | 33 | 7,092 |
| Median F0 | Outside 110–155 Hz | Interjections: off | 146 | 7,271 |
| UTMOS | Below 2.8 | Up to 4 words: off | 4,832 | 921 |
| System | Weights | WER, % | CER, % | Clean, % | UTMOS | Parameters |
|---|---|---|---|---|---|---|
| Public harness, 495 sentences, 3 takes | ||||||
| EMA Lightning 1.0.1 | open | 1.04 | 0.225 | 93.7 | 3.30 | 8.6M |
| Antalia-2 Mini 1.0.0, previous release | open | 1.05 | 0.200 | 93.1 | 3.25 | 7.62M |
| Antalia-2 Mini 1.0.1 | open | 1.09 | 0.220 | 92.9 | 3.18 | 7.62M |
| Alania-2 API, voice Öncel (12 Oct) | closed | 1.15 | 0.217 | 93.0 | 3.52 | 488M a |
| Trendyol-TTS | open | 1.16 | 0.229 | 93.1 | 3.82 | 2.38B d |
| Mini 1.0.1 minus | WER | CER |
|---|---|---|
| EMA Lightning 1.0.1, public harness | +0.05 [ 0.22, +0.32] | 0.005 [ 0.075, +0.065] |
| EMA Lightning 1.0.1, own harness | +0.02 [ 0.25, +0.27] | +0.007 [ 0.052, +0.071] |
| EMA Lightning 1.0.1, pooled over 6 takes each | +0.03 [ 0.21, +0.28] | +0.001 [ 0.057, +0.062] |
| VoxCPM2 teacher, own harness, 1 seed | +0.10 [ 0.25, +0.45] | +0.005 [ 0.103, +0.105] |
| File | Bytes | Holds |
| Weights, fp32, default download | 30,499,764 | Acoustic 14,757,148; vocoder 15,742,616 |
| Weights, fp16, variant="fp16" | 15,257,774 | Acoustic 7,383,130; vocoder 7,874,644 |
| Python wheel, 1.0.1 | 1,397,120 | Inference code and normalizer data, no weights |
| Browser graphs, mixed, page default | 22,848,212 | Text, condition, flow in fp32; vocoder in fp16 storage |
| WebAssembly (WASM) binary | 14,239,897 | ONNX Runtime Web 1.30.0 |
| Android APK, arm64-v8a | 35,549,175 | Engine v0.1.1, model 1.0.1 |
| Public | Not public |
|---|---|
| Weights, fp32 and fp16 (Hugging Face) | Training code |
| Inference code, Python normalizer (PyPI wheel, sdist) | Data generation scripts |
| JavaScript normalizer, browser page, ONNX graphs (Space) | Training data, the chosen reference clip |
| Benchmark harness, adapters, per-clip scores (Hugging Face) | Own-harness code |
| Signed Android APK, binary only | ONNX export and verify tools, parity tests |
| Hosted API, needs a key | Android source, Kotlin normalizer source |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
TTS Synthesis
Text-to-speech synthesis converts written language into intelligible, natural-sounding speech with control over voice, prosody, style, and timing.
See all 294 papers