> ## Documentation Index
> Fetch the complete documentation index at: https://docs.slng.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Text-to-Speech Overview

SLNG gives you access to a range of text-to-speech models from multiple providers through a single platform. The execution layer assembles output from cache when possible, so repeated phrases return from the edge without hitting the upstream model.

For the full, current list with latency, hosting, and language details, see [Text-to-Speech Models](/models/tts). For voice samples, see [Voices](/models/index).

## Which approach should you use?

| I need to...                                        | Use                                                           |
| --------------------------------------------------- | ------------------------------------------------------------- |
| Generate audio files for download or storage        | [HTTP](/examples/tts-http)                                    |
| Stream audio in real time for a voice agent or app  | [WebSocket Streaming](/examples/tts-websocket)                |
| Control how brand names and acronyms are pronounced | Add [Pronunciation Dictionaries](/pronunciation-dictionaries) |
| Use my own provider keys with SLNG's optimization   | Enable [BYOK](/execution-layer/byok)                          |

## Which model should you use?

{/* TODO(content): verify these model recommendations against models/models.json before publish — names and language counts must match the catalog. */}

| I need...                                   | Recommended model                     | Why                                          |
| ------------------------------------------- | ------------------------------------- | -------------------------------------------- |
| Lowest latency English voice agent          | Deepgram Aura 2 (SLNG-hosted)         | Low-latency, deployed on SLNG infrastructure |
| Lowest latency with broad language coverage | Cartesia Sonic 3                      | Many languages, low latency                  |
| Hindi / Indian languages                    | Sarvam Bulbul                         | Indian languages, many voices                |
| Multilingual with expressiveness control    | KugelAudio Kugel                      | Broad language coverage                      |
| Spanish                                     | Deepgram Aura 2 Spanish (SLNG-hosted) | SLNG-hosted Spanish voice                    |

## Execution layer behavior

Every TTS request goes through the output assembly stage, which checks whether this text, voice, and config combination has been assembled before:

* **Cache hit**: audio returns from the edge. No upstream model call. No provider billing.
* **Cache miss**: the request routes to the model endpoint, audio is synthesized, and the result is cached.

The more calls flow through the system, the more cache coverage grows. Repeated phrases (greetings, disclosures, confirmations) become effectively free.

See [Output Assembly](/execution-layer/output-assembly) for caching, segment reuse, and cache scoping.

## Quick example

```bash theme={null}
curl https://api.slng.ai/v1/tts/slng/deepgram/aura:2 \
  -H "Authorization: Bearer SLNG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello from SLNG!"}' \
  --output hello.wav
```

See [HTTP](/examples/tts-http) for complete examples in cURL, JavaScript, and Python, and [WebSocket Streaming](/examples/tts-websocket) for real-time voice agent audio.