In Retell, there isn’t a true “train from uploaded call recordings” feature in the sense of the agent learning directly from audio files inside the UI. What you can do instead is use call transcripts—once calls are recorded and transcribed, you can copy those transcripts and use them to improve your prompt, knowledge base, or few-shot examples manually. So the “training” is really prompt refinement rather than automatic learning from uploads. For supporting languages that aren’t in the default list, the best approach is to rely on the underlying model’s multilingual ability and explicitly add that language into your prompt instructions and test it with examples. If it’s a less common language, quality will depend heavily on the model, so you may also want to validate it through your workflow layer (like detecting language first and routing appropriately) to keep responses consistent. @jsdev