canberk7 / ema-lightning

# ⚡ EMA Lightning
**Tiny, fast and accurate Turkish text to speech.** 8.6M parameters, about 34 MB,
Apache-2.0. It runs on your own GPU or CPU, offline: no text or audio leaves the
machine.
- **The most accurate on Freya-TR-Eval:** 0.92% word error rate, the lowest of every
 system measured, including ElevenLabs v4, Gemini 3.8 and the 2.38B-parameter
 Trendyol-TTS.
- **Fast:** the first audio is ready in about 4 ms. One request runs 440× faster
 than real time, and batched work 1,316× (RTX 4090).
- **Cheap:** about $0.0085 per million characters on a rented RTX 4090.
- **Takes Turkish as people write it:** numbers, dates, times, money, units,
 abbreviations and symbols are read aloud.
- **Streams:** audio starts while the rest of the text is still being made.
- **One model serves many callers:** call it from as many threads as you like.
 Its built-in scheduler, Playhead, runs everyone's work together on the GPU. No
 server or queue system to run.
| System | Size | Word error rate | Speed, RTX 4090 |
|---|---|---|---|
| **EMA Lightning** | **8.6M** | **0.92%** | **440× real time** |
| Trendyol-TTS | 2.38B | 0.98% | 2.8× |
| Gemini 3.8 Flash-Lite | cloud | 1.31% | — |
| ElevenLabs v4 | cloud | 1.43% | — |
| Anka TTS † | 336M | 1.73% | 6.3× |
| XTTS-v2 † | 470M | 3.34% | 7.1× |
| FreyaTTS † | 183M | 12.02% | 8.4× |
Freya-TR-Eval, all 495 sentences, speed 1.0. † Reported on the Anka TTS model card.
The full table and how it was measured are on the
model card.
## Install
 [code] 
CPython 3.11 to 3.13 with PyTorch 2.1 or newer. The weights (about 34 MB) download
from Hugging Face on first use.
On Linux the default PyTorch build includes CUDA, which pulls in about 1.5 GB of
NVIDIA packages. On a machine without an NVIDIA GPU, install the CPU-only build
instead:
 [code] 
Thanks to @keyiflerolsun for spotting the 1.5 GB
of unneeded NVIDIA packages on CPU machines
(#4).
## Use
Make one `EMA` and keep it for the whole program. There are two ways to get
speech out of it:
| | `say()` | `stream()` |
|---|---|---|
| Gives you | the whole audio, once it's ready | the audio piece by piece, as it's made |
| Takes | one text, or a list of texts | one text per call |
| Use it for | files, voiceovers, batch jobs | playing as you go: speakers, phone calls, live apps |
### `say()`: a whole text, back when it's ready
 [code] 
`say()` returns a `Speech`: the `audio` (a float32 NumPy array, mono, between -1
and 1), its `sample_rate`, its `duration` in seconds, and the `seed` that made it.
With `path`, it also writes a WAV file. Pass the same `seed` again to get the same
audio.
Give it a list and every text is made in shared batches, one `Speech` per text, in
order. With `path`, the list is written into that folder as `0.wav`, `1.wav`, …:
 [code] 
### Word timings: when each word is heard
Every `Speech` from `say()` carries `words`: each spoken word with its `start` and
`end` in seconds. They come from the same frame plan the audio is made from, so
they cost nothing extra