Can skipping transcription make voice assistants faster?

Conventional voice assistants run a three-stage pipeline: automatic speech recognition (ASR) converts speech to text, an LLM generates a text response, and text-to-speech (TTS) renders that response as audio. Each stage adds latency and propagates errors, and the total response time is dominated by stage-by-stage processing rather than by the LLM's reasoning. LLaMA-Omni eliminates the transcription step entirely. It integrates a pretrained speech encoder, a speech adaptor, an LLM (built on Llama-3.1-8B-Instruct), and a streaming speech decoder so that the system generates text and speech responses directly from speech instructions, achieving a response latency as low as 226 milliseconds.

The architectural lesson is that transcription is not a free intermediate representation — it is a serialization step that destroys prosodic information and forces full-utterance processing before generation can begin. By passing speech embeddings directly into the LLM via an adaptor, LLaMA-Omni preserves the acoustic information the LLM might use — including the articulatory substrate identified in Do speech models learn language-specific sounds or universal physics? — and lets generation begin as soon as enough of the input has been encoded, without waiting for a complete textual transcript.

The supporting piece is the InstructS2S-200K dataset of 200K speech instructions paired with speech responses, which makes the alignment trainable. Without paired speech-to-speech instruction data, end-to-end training has nothing to optimize. The general principle: when latency dominates user experience, the right intervention is to remove pipeline stages, not to optimize each stage in isolation.