Flow Head

A flow head is a flow-based neural network decoder that generates audio output in end-to-end audio-generation models. In a conventional text-to-speech pipeline, a language model produces text or acoustic features, which are then passed to a separate neural vocoder (often a separate flow or diffusion model) to generate the actual waveform. In the interaction model's encoder-free early fusion architecture, audio generation is handled by a flow head — a small, specialized flow-based network that is co-trained end-to-end with the core transformer, rather than being a separately pre-trained component.

Flow-based models (normalizing flows) learn an invertible transformation between a simple distribution (typically Gaussian noise) and a complex data distribution (in this case, audio waveforms or dMel features). They are trained by maximizing the exact log-likelihood of the data, which makes them computationally tractable and amenable to end-to-end training. In the audio generation context, the flow head takes the transformer's output representations and transforms them into the audio domain — producing the actual acoustic signal that the user hears as speech.

The advantage of co-training the flow head with the rest of the system is that the generation objective is aligned with the reasoning objective from the start of training. A separately trained vocoder may prioritize acoustic fidelity (making speech sound natural and clear) at the expense of semantic alignment (producing the right words in the right order at the right time). When the flow head is part of the end-to-end objective, the model can optimize directly for the full generation quality — including semantic content, prosody, timing, and naturalness — rather than optimizing for acoustic quality alone and hoping that semantic alignment emerges as a side effect.

The use of flow-based generation rather than autoregressive or diffusion-based alternatives reflects a balance between quality and latency. Flows are typically faster to sample from than diffusion models and can produce high-quality audio with fewer network evaluations. The tradeoff is that flows require carefully designed architectures to be expressive enough for complex data distributions, and training stability can be a challenge for high-dimensional audio data.