An easy example of this issue is the degradation of the transcript / model input itself. LLM output: 明白了。请您详细描述一下脚痛的情况吗？例如，痛了多久了，痛在哪里？ => Understood. Could you please describe your foot pain in detail? For example, how long has it been hurting and where is it? Transcript record after TTS -> STT: 明白了,请您详细描述一下交通的情况吗?例如,通了多久了?通在哪里? => Got it, could you please describe the traffic situation in detail? For example, how long has it been on the road? Where is it? The above was using 11labs and deep gram, but we’ve seen these issues on other combinations eg azure and talkscriber. Another more interesting example is the transcript being out of sync. In this scenario we have a relatively long opening message in simplified cheese. The user responds with a short phrase. Vapi then sends us a transcript / model input in the order: - System prompt - Short user message - Assistant intro message. This is never rectified and even the end of call summary is out of order. If the assistants turn is run through STT, I can only assume that this is causing a race condition where the user’s message is parsed before the assistant message which then corrupts the transcript. Can someone please confirm if Vapi is using STT on the voice model’s output for the call transcript? If so, is there some way we can disable this to make the LLM output the source of truth? If this isn’t the cause I would be keen to hear some people’s thoughts on what could be going wrong.