A voice-and-image support flow
- The customer speaks and STT creates a transcript.
- The customer attaches a photo and the VLM describes relevant visible evidence.
- The LLM combines the transcript, image findings, policy, and account data.
- A human approves a high-risk decision.
- TTS reads the approved answer back to the customer when voice output is useful.
Keep the boundaries clear
Do not ask one model to hide all the pipeline’s errors. Store the transcript, image interpretation, retrieved evidence, decision, and final speech separately. When the answer is wrong, you need to know whether the audio was transcribed incorrectly, the image was misunderstood, the policy was missing, or the LLM reasoned badly.
The first version should be smaller
For a voice-to-content product, start with STT plus an LLM and a human review step. Add TTS when people actually need audio output. Add VLM only when visual input changes the task, not because a multimodal architecture sounds more complete.