Hey everyone, ​ ​Like many of you, I’m looking into the newly released Gemma 4 12B to build a native speech-to-speech experience. Because of its unique encoder-free architecture, completely skipping the traditional STT bottleneck could be possible. ​ ​Right now, my main focus is strictly on the input side: I want a low-latency, native voice ingestion workflow without writing a massive, complex pipeline from scratch. ​ ​Are there any reliable solutions that fully ...