Antalia-2 Mini: a Turkish TTS model with 7.6 million parameters trained only on synthetic speech

A team of researchers has released Antalia-2 Mini, a single-voice Turkish text-to-speech model with 7 620 931 parameters. The model reads text character by character and generates audio at a sampling rate of 48 kHz at a speed of 41 times real time on the CPU of an Apple M5 laptop (2 threads). The training audio came exclusively from the VoxCPM2 model acting as a "teacher," which is roughly 300 times larger — after automatic quality checks, approximately 1 790 hours of synthetic data remained.

Fine-tuning the vocoder on the acoustic model's own mel spectrograms reduced the Whisper transcription error rate (WER) from an average of 3.89% to 1.75% between adjacent preliminary checkpoints. The authors also mention a curiosity: the teacher model corrupts the beginning of a clip whenever the first word begins with the letter "ö," and the student learned to replicate this error, which the authors hide in the output by trimming the prefix.

On the Freya-TR-Eval benchmark (495 sentences), on which the model's parameters were tuned, version 1.0.1 achieves a WER of 1.09% and a character error rate (CER) of 0.22%. A paired bootstrap test did not distinguish the results from the EMA Lightning model (8.6 million parameters, also Turkish), which however scores better on the UTMOS metric across the full frequency spectrum — though this metric is trained on English. According to the authors, no test with human listeners exists.

The weights, inference code, and benchmarking tool are publicly available under the Apache-2.0 license; the training code and data have not been released.