> ## Documentation Index
> Fetch the complete documentation index at: https://docs.geko.sh/llms.txt
> Use this file to discover all available pages before exploring further.
# Latency & quality (NFE)
> Tune the nfe dial and understand cold starts.
Two things dominate how fast you get audio and how good it sounds: the **`nfe`** setting and **cold starts**.
## `nfe` — the quality/speed dial
`nfe` (number of function evaluations) is how many diffusion steps the model runs. More steps = better quality, more time.
| `nfe` | Use it for |
| -------------- | ------------------------------------------------------------------- |
| `16` | Low-latency / interactive — noticeably faster, very usable quality. |
| `32` (default) | Best quality — narration, published audio, anything you'll ship. |
```ts theme={null}
await geko.tts.create({ text, voice: "Aigerim", nfe: 16 }); // snappy
await geko.tts.create({ text, voice: "Aigerim", nfe: 32 }); // polished
```
Rule of thumb: start at `32`; drop to `16` when responsiveness matters more than the last bit of fidelity.
## Cold starts
The GPU backend **scales to zero when idle** to save cost. Consequences:
* The **first** request after an idle period pays a cold start — this can be tens of seconds.
* After a request, the service stays **warm for a few minutes**, so follow-up calls are fast.
* The SDK's default **120 s** timeout is sized for this. Don't lower it below what a cold start needs unless you know the service is warm.
**Keeping it warm:** if you have predictable traffic, a lightweight periodic `geko.health()` (or a pre-warm call before a known burst) keeps a container hot. Always-on warmth is a server-side deployment choice.
## Other factors
* **Text length.** Longer text = longer synthesis (roughly linear). Break very long inputs into sentences/paragraphs and synthesize in parallel if you need throughput.
* **`speed`.** Changes playback pace (0.5–2), not generation time.
* **Concurrency.** Independent requests run concurrently; the SDK doesn't serialize them.
## Measuring
Time a warm call end-to-end:
```ts theme={null}
const t = performance.now();
await geko.tts.create({ text: "Сәлем", voice: "Aigerim", nfe: 16 });
console.log(`${Math.round(performance.now() - t)}ms`);
```
Run it twice — the second number (warm) is the one your users will usually see.
## Next steps
Start playback sooner by synthesizing sentence-by-sentence.
Wire low-latency TTS into a real-time voice loop.
Text limits, output format, and cold starts.