EMA Lightning publishes weights, code and author-run benchmarks for an 8.6M-parameter streaming Turkish TTS stack.
EMA Lightning releases a compact streaming Turkish speech model
2026-10-01Frontier watchTRIAL
EMA Lightning releases Apache 2.0 weights and code for an 8.6M-parameter Turkish text-to-speech stack with streaming, CPU support and a shared GPU scheduler. The model card reports 0.92% WER on Freya-TR-Eval, 3.86 ms to first audio and 440x real-time generation on an RTX 4090. Training used about 1,000 hours of private single-speaker audio.
What this means for you
For teams serving Turkish speech, this is a small, inspectable alternative with unusually low reported latency. The benchmark and cost numbers are author-run, and the private training corpus limits data-provenance review, so production use needs independent quality and safety testing.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Parameters | 5.6M acoustic model plus 3M decoder; about 34 MB total |
|---|---|
| License | Apache 2.0 |
| Language | Turkish |
| Reported Quality | 0.92% WER on Freya-TR-Eval |
| Reported Speed | 3.86 ms first audio and 440x real time on RTX 4090; about 6x real time on CPU |
| Training Data | about 1,000 hours of private single-speaker Turkish speech |
| Artifacts | Hugging Face weights, GitHub code and pip package |
Technical impact 3
Engineering value 4
Adoption signal 3
Maturity 3
Verification cost 2
Risk 3