@AI Evy Hi, I know that pronunciation dictionary is only available with English & 11labs. However my use case is non-English and I need to somehow change how the agent pronounces certain words. What are my options here?

Thanks for tagging me @Ozan. I'll look through relevant sources and get back to you shortly.

@AI Evy Yes, provide a general answer on your broader knowledge. Maybe configure the LLM part to output certain words in a way that we want how them to be pronounced? (like regex or basic text replacement?)

### Answer Yes, that's a viable workaround. Here are some general approaches you can try: ## 1. Prompt-Based Phonetic Spelling Instruct your LLM (via the global prompt or node-level prompts) to output certain words spelled phonetically in a way that the TTS engine will pronounce correctly. For example, if "Müller" is mispronounced, you could instruct the LLM: "When saying the name Müller, output it as 'Mew-ler' instead." ## 2. Custom LLM Post-Processing If you're using a Custom LLM via the LLM WebSocket, you have full control over the

content

field in your response events. [(1) ]() You can implement text replacement logic on your server before sending the response:

python
# Example server-side replacement before sending to Retell
content = content.replace("Müller", "Mew-ler")
content = content.replace("Szczepański", "Shche-pan-ski")

This gives you regex or dictionary-based substitution capabilities since you control what text gets sent to the TTS. ## 3. Speech Normalization Setting Retell offers a Speech Normalization option that converts entities like dates, currency, and numbers into plain words, which can help with pronunciation issues for those specific cases. [(2) ]() ## Limitations These are workarounds rather than native pronunciation dictionary support. The phonetic spelling approach works best when: - You have a finite set of problematic words - The phonetic approximation sounds acceptable in your target language - You're willing to accept slightly unnatural spellings in transcripts For production use cases with many pronunciation requirements, reaching out to Retell support about non-English pronunciation roadmap would be worthwhile.