streaming as a first-class citizen. The first step is to completely rethink the encoder. Toss away the bidirectional attention and replace it with causal attention so that each audio chunk only depends on previous ones. In theory, this should be enough to enable streaming behavior. But as I mentioned earlier, the biggest value proposition of DSM is to enable pre-existing, pre-trained LLMs to bootstrap speech models. Since most LLMs today are shaped as a transformer and decoder, the decoder is a natural place to plug them in. In particular, Vox Real Time leverages Mistral 3B. This is a completely different LLM from the one that sits in the middle of a modular voice assistant. It's an internal component of the speech-to-text model and it gets fine-tuned for this task. Given the presence of this LLM, we also need to reconsider how the encoder connects to the decoder, since this cross-attention doesn't come with the LLM. Instead, we'll move the IO encoder downstream. Once contextualized, audio embeddings are directly added to the corresponding token embeddings. This way, audio and text are processed at the same pace by design. From the point of view of the LLM, there's a single input token for each heartbeat. It's just that under the wraps, that input token is the sum of audio and text embeddings. And this is the architecture of real-time speech-to-text models that follow the DSM framework. This is what Boxcar Realtime builds upon, making additional adjustments that ensure training stability. One of their innovations is a technique called adaptive RMS norm. It's a new way of communicating to the model the desired delay selected by the user at inference time. If you're familiar with such architectural details, you'll immediately think of adaptive layer norm, a similar trick used by the diffusion transformer to inject the diffusion step, which I've covered in this other video. So, once audio is transcribed by a