FUTURE WORK / MODEL

MacroCF

Autoregressive audio–text modeling for Conversation Flow.

MacroCF is a planned model that would condition conversational decisions on the acoustic stream, text context, and dialogue history. The goal is to distinguish situations that may sound similar but have different conversational meanings: a completed request, an unfinished thought, an acknowledgement, or speech that does not call for a response.

A decision every 200 ms

We target one decision update every 200 ms (5 Hz). Each update would use the accumulated context available so far. This is a decision-frequency target, not a 200 ms context window or a measured end-to-end inference latency. Decisions would remain causal.

Structured conversational decisions

The proposed output vocabulary includes five control tags:

<EOT/>   <INT/>   <LTN/>   <BC/>   <REJ/>

These express turn completion, interruption, continued listening, backchanneling, and deliberate non-response. Their semantics and examples are documented in the event interface.

Optional reasoning output

A runtime option is planned to enable or disable an explicit, brief rationale in a <think> block. With the option enabled, a decision could be expressed as:

<think>The speaker has completed a request and is yielding the floor.</think>
<EOT/>

With it disabled, the output would contain only the control tag. This switch describes the proposed output interface; any effect on computational cost, latency, or decision quality must be measured.

Data and evaluation direction

MicroCF could support candidate selection and annotation prioritization for data development. MacroCF would be evaluated on the contribution of acoustic and semantic context, behavior under matched false-positive budgets, and reasoning-enabled versus direct-tag output. Candidate labels would be reviewed, and held-out evaluation data would remain separate.

MacroCF is future work. MicroCF’s published dev results do not establish performance for MacroCF or for all five proposed tags.