Voice Agent Stack Evolves: Ditching STT for Direct Multimodal

andimarafioti · x · 2026-08-11

Traditional voice assistants rely on a cumbersome VAD -> STT -> LLM -> TTS pipeline. Now, developers can leverage the native audio understanding capabilities of Multimodal LLMs (MLLM).

The new minimalist stack simplifies this to VAD -> MLLM -> TTS, completely removing the Speech-to-Text (STT) step. The model listens to and processes audio directly, which will significantly reduce latency and streamline the development of voice agents.

Original post →

More from coding & agent

coding & agent channel →