Mechanism: An included Turkish text normalizer handles numbers, dates, times, amounts, abbreviations, URLs, and English or brand terms before synthesis; package, streaming, and browser-local paths provide routes to integrate or test the model.
Antalia-2 Mini has 7.62 million parameters and produces 48 kHz audio; the model card says it runs faster than real time on a CPU. It provides a Python package, streaming and batched synthesis, and a browser demo that uses ONNX Runtime Web locally. The weights and code use Apache-2.0.
Why now: The release packages language-specific normalization, local execution options, and permissive licensing in a compact model instead of requiring a hosted endpoint. The model card reports comparisons with EMA Lightning, including similar WER in its harness, but also differing naturalness scores across measures and bands. Those results are useful as a starting point, not proof of performance across applications or users.
First implication: Builders can test Turkish speech generation in CPU-based or local workflows without first relying on a hosted service.
The model card describes one synthetic male voice and says its training data contains no recordings of real speakers. That narrows one potential concern, but does not establish broad voice coverage, robustness, or suitability for every deployment. Evidence boundary: Release specifications and evaluation results are taken from the project's Hugging Face model card and have not been independently verified.
First watch condition: Independent tests should establish how intelligibility and naturalness hold across varied Turkish text, devices, and real application audio conditions.
This starts a distinct analytical record because no prior New Runtime record was supplied for this specific Turkish TTS release; no related record is asserted.