for Turkish with Eleven labs, pronunciation dictionaries aren’t available yet. The most reliable approach is text-level control: adjust spellings to be more phonetic in Turkish, use SSML where Eleven lab s supports it, or preprocess the text (string replacements) before sending it to Retell. This usually works well for fixing specific names or recurring words. @Ozan

yeah I was thinking about the same; but we will be using the full pipeline of STT - Reasoning - TTS obviously; so I dont think I would be able to send an already preprocessed text to TTS; unless we make a custom LLM and intervene with websockets. thats one way; the other way would be "fixing" the text replacements in the LLM level (via prompts) I guess. which would not grant full determinism but would probably work for most of the cases

You’re thinking in the right direction. With the full STT → LLM → TTS pipeline, direct TTS preprocessing isn’t really available unless you intercept the stream (e.g. via a custom LLM + websocket layer). In practice, most teams handle this at the LLM level: prompt-based text replacements or a lightweight post-LLM normalization step before TTS. It’s not 100% deterministic, but for Turkish it usually works well for names and common terms. If you need stronger guarantees, the websocket interception approach is the only truly deterministic option right now. @Ozan