Vapi performs a filter for low confidence words or phrases received from the Transcriber (like Deepgram). Often times, when there are foreign accents, or foreign names, the transcriber transcribes something (probably wrong), and Vapi filters out every single word, so nothing gets sent to the model. This often results in a silence timeout for the call. It would be great if we could set the confidence threshold for filtering out words, so that at least the models gets something, even if wrong.