Pronunciation dictionaries

A pronunciation dictionary is a reusable set of rewrite rules. Each rule swaps a word or phrase for replacement text before it reaches the TTS model, so “SLNG” is spoken as “slang” and “NAIC” as “en ay eye see” on every turn. You create the dictionary once, then reference it by name when you open a TTS session. Rewrite matching is deterministic: case-insensitive, whole-word or whole-phrase, longest-match first, and non-recursive. The walkthrough below uses a WebSocket, but the same pronunciation selector applies to HTTP TTS requests. One dictionary is active per request or WebSocket turn.

Prerequisites

An SLNG API key.
The hostname of a production cluster with dictionary support enabled.

Create and consume a dictionary on the same cluster, organization, and environment. A dictionary created on one cluster is not visible from another.

Create your dictionary

Send the rewrite rules to POST /v1/pronunciation/dictionaries. A successful call returns 201 Created.

export SLNG_API_KEY="your-api-key"
export SLNG_BASE_URL="https://[region].api.slng.ai"

curl --fail-with-body "$SLNG_BASE_URL/v1/pronunciation/dictionaries" \
  -H "Authorization: Bearer $SLNG_API_KEY" \
  -H "Content-Type: application/json" \
  --data '{
    "name": "brand-names",
    "modes": {
      "rewrite": {
        "rules": [
          {"match": "SLNG", "replace": "slang"}
        ]
      }
    }
  }'

Each rule pairs a match term with the replace text sent to the model. A dictionary holds up to 256 rules. Pronunciation fails closed: if a request names a dictionary that cannot be resolved, the TTS request is rejected rather than synthesized without it.

Open a TTS WebSocket with the dictionary selected

Connect to a TTS endpoint on the same cluster and pass the dictionary name in the init config. This example uses Deepgram Aura 2:

wss://[region].api.slng.ai/v1/bridges/unmute/tts/deepgram/aura:2

Send the API key in the handshake header:

Authorization: Bearer your-api-key

Then send the init message with a pronunciation selector:

{
  "type": "init",
  "config": {
    "encoding": "linear16",
    "sample_rate": 24000,
    "pronunciation": {"mode": "rewrite", "name": "brand-names"}
  }
}

Send text and collect audio

After the server sends ready, send a text message:

{"type": "text", "text": "Welcome to SLNG.", "flush": true}

Collect binary audio frames until audio_end. The model speaks “SLNG” as “slang” instead of reading the letters on its own.

Reuse the selection across turns

The selection stays active for the session. Send later text messages without repeating pronunciation. To switch dictionaries, include a new selector on the first text message of a later turn.

Keep complete matching words or phrases within a single message, and set the dictionary before a turn begins rather than mid-turn. To speak without rewriting, open a new session with no pronunciation selector.

API reference