korhun/ema-lightning-onnx

EMA Lightning — ONNX

ONNX export of EMA Lightning by Canberk Aslan — a natural Turkish text-to-speech model (8.6M parameters, Apache-2.0) — so it runs on ONNX Runtime, offline, without Python or PyTorch (for example from .NET).

All credit for the model goes to Canberk Aslan. This repository only converts the published weights into three ONNX graphs and checks them against PyTorch. Numbers, dates, money and units are turned into words by normalizer-tr by Erdem Tuna (Apache-2.0), which runs before the model.

File	What
text.onnx	letters → features and per-letter durations
sound.onnx	features + frame plan + noise → latents (the four distilled steps unrolled)
decoder.onnx	25 Hz latents → 48 kHz audio
vocab.json, model-config.json	the model's alphabet and constants

Exported from revision 7a6ba1ad of canberkkkkkk/ema-lightning with export.py: largest audio difference against PyTorch 5e-4, identical frame plan. On a desktop CPU it runs about 20× faster than real time (first sentence in 0.3–0.5 s, ~230 MB memory).

Use it

• .NET: ema-lightning-dotnet — the frontend, chunking, frame plan and vocoder loop on ONNX Runtime, plus a say sample.
• Windows app: Sesli Okuma — reads selected text aloud in any app with this voice, fully offline; free and open source (winget install KorayOrhun.SesliOkuma).
• Browser demo: Türkçe Doğal Ses.

Responsible use

Tell listeners the audio is AI-generated, and do not use the voice for impersonation, fraud or misleading content (from the original model card).

License

Apache-2.0, like the original model — see LICENSE and NOTICE. Changes: the PyTorch weights were converted to ONNX graphs; nothing was retrained.