Antalia-2 Mini ships a 7.6M-parameter Turkish voice trained only on a teacher about 300 times larger
Synopsis
Antalia-2 Mini is a 7,620,931-parameter single-voice Turkish text-to-speech model whose training audio was generated entirely by VoxCPM2, a teacher roughly 300 times larger, cloning one synthetic reference clip, with about 1,790 hours kept after automatic checks. On the 495 sentences of Freya-TR-Eval it reaches 1.09% WER and 0.220% CER, a paired bootstrap cannot separate it from the 8.6M EMA Lightning, the authors tuned settings on those sentences, and no human listening test exists.
Interpretation
A 7.62M-parameter single-voice Turkish model synthesizes 48 kHz audio at 41 times real time on a laptop CPU, with 30.5 MB default weights and an optional 15.3 MB fp16 download. Two small open Turkish models run earlier, Piper (voice dfki, 16M) and MMS-TTS (36M), have word error rates of 3.41% and 6.38%, against 1.16% for Trendyol-TTS at about 2.38B. On the 495 sentences of Freya-TR-Eval with the public harness, version 1.0.1 scores 1.09% WER and 0.220% CER, and paired 95% bootstrap intervals cannot distinguish it from EMA Lightning 1.0.1 or, on the authors' own harness, from the teacher at one seed.
The model reads characters with no phonemizer, and a text normalizer first writes numbers, dates, initialisms and English words into the 47 symbols it knows. An acoustic model maps characters to mel spectrograms, a generator trained with shortcut self-consistency samples in 1 to 8 steps, and a vocoder turns the mel into a 48 kHz waveform in one pass. At default settings the generator does 84.0% of the multiply-accumulates, a 5.11 s sentence costs 14.49 GMAC, and on two threads of an Apple M5 the first 100 sentences run at a median 41 times real time.
The vocoder is fine-tuned on mel spectrograms the acoustic model generates, not only on real mels. A vocoder trained only on real mels left word error rate above 3.6%, and a frozen vocoder stayed at 3.63% to 4.14%. Across 7 pairs of adjacent pre-release checkpoints at one seed and 4 sampling steps, the fine-tune cut mean WER from 3.89% to 1.75%, with drops of 1.66 to 2.63 points per pair and every paired interval excluding zero.
Perspective
The demonstrated scope is single-voice, single-style Turkish-only synthesis, scored on the 495 sentences of Freya-TR-Eval with the authors' harness; the browser page and signed Android engine were checked on an emulator, and no real phone has been measured.
The authors state that no human listening test exists, that final-settings numbers are optimistic because settings were tuned on the same sentences, and that UTMOS is an English-trained proxy compared across two different voices; the teacher's defect on ö-initial words was learned by the student and is hidden, not removed, by a spoken prefix cut from the output.