PersonaKit: Configurable Dialogue Personas
- The paper introduces PersonaKit, an open-source framework that treats interruption behavior as a key component of persona design.
- It separates low-latency audio processing from JSON-configured experimental controls, enabling flexible role-specific dialogue management.
- Pilot evaluations demonstrate that tailored turn-taking policies enhance perceived realism and consistency in diverse spoken agent personas.
Searching arXiv for the target paper and a few related persona/dialogue works for accurate citation support. PersonaKit (PK) is an open-source, plug-and-play platform for building and evaluating full-duplex spoken dialogue agents whose interruption behavior is treated as part of persona design rather than as a fixed backend default. It is motivated by the claim that, as spoken agents move beyond generic assistant roles into socially specific characters such as “authoritative instructors, uncooperative merchants, or distracted workers,” the quality of interaction depends not only on semantic content but also on how the agent manages overlapping speech. In this view, yielding, persisting, bridging, or overriding during barge-in are socially meaningful acts that shape dominance, cooperativeness, authority, attentiveness, and psychological immersion. PersonaKit operationalizes this claim through a low-latency web platform, JSON-based persona and interruption policies, live deployment tooling, and automatic post-session survey workflows for comparative user studies (Jeon et al., 7 May 2026).
1. Conceptual basis
The central premise of PersonaKit is that turn-taking in spoken interaction is not a neutral transport-layer behavior. In human conversation, whether a speaker yields, resumes, bridges, or talks over an interruption conveys social stance. The paper therefore argues that the common full-duplex “always-yield” policy—immediately ceding the floor whenever user speech is detected—may be serviceable for submissive assistant-like roles but can break immersion for non-submissive, inattentive, or institutionally authoritative personas (Jeon et al., 7 May 2026).
The motivating examples are deliberately role-specific. A military drill sergeant who instantly stops speaking whenever the user barges in becomes psychologically implausible. Likewise, a tavern keeper, salesperson, DMV clerk, or distracted chef should not all respond to interruption identically. PersonaKit reframes overlap handling as a first-class sociolinguistic design variable, rather than a backend constant.
This positioning differentiates PersonaKit from other persona-oriented systems that act on different layers of agent behavior. COSPLAY models both self persona and partner persona in text dialogue through concept-set operations and future-dialogue rewards (Xu et al., 2022). MASK embeds persona into an embodied robot through a finite-state machine over motion-expression states (Park et al., 2024). RoleMemo and DualMem focus on persona-conditioned memory and long-term interpretive consistency in role-playing agents (Zhang et al., 25 May 2026). PersonaKit instead targets the floor-management layer of full-duplex spoken interaction: interruption handling itself becomes the persona-bearing mechanism (Jeon et al., 7 May 2026).
2. System architecture and configuration
Architecturally, PersonaKit separates low-latency audio mechanics from experimental design. The stack is Python/Flask on the server with Socket.IO and vanilla JavaScript in the browser. Researchers operate the system through a web dashboard and a small set of JSON files rather than by modifying source code, and the paper stresses that “all experimental behavior is configured through JSON—no source code is modified” (Jeon et al., 7 May 2026).
The main configuration artifacts are the following:
| File | Function |
|---|---|
persona.json |
scenario, role, opening prompt |
interruption_config.json |
strategy matrix for interruption handling |
session_config.json |
post-session survey configuration |
model_config.json |
LLM/TTS routing and provider selection |
This JSON-driven design is the basis of the platform’s plug-and-play character. Persona description, opening prompt, role, scenario, and interruption behavior can all be altered without changing code. The same backend and even the same underlying LLM can therefore be reused while only changing policy weights. The paper’s demonstration contrasts a “Grumpy Tavern Keeper” configured to hold the floor with a “Standard AI Assistant” configured to yield, under the same general infrastructure (Jeon et al., 7 May 2026).
On the client side, the browser uses WebRTC microphone capture. A client-side VAD “volume gate” rapidly detects user speech and immediately halts local playback on barge-in. This is a crucial latency optimization because stopping playback in the browser avoids waiting for a server round trip before the user perceives that the interruption has been registered. PersonaKit also tracks playback progress at byte level and records exactly which portion of the bot’s TTS output had already been vocalized. It logs both the “cutoff text” and the “remaining text,” which are then sent to the server so the dialogue manager and LLM know the exact interruption point and can recover coherently (Jeon et al., 7 May 2026).
On the server side, the interrupted utterance is transcribed by ASR, passed to a zero-shot intent classifier, routed through a turn-taking manager, then sent to generation and TTS. Under the authors’ OpenAI plus ElevenLabs configuration, the reported default end-to-end barge-in latency is approximately ~1--2 s (Jeon et al., 7 May 2026).
3. Interruption modeling and prompt control
PersonaKit’s interruption model is explicit but intentionally lightweight. User interruptions are classified into four intent classes grounded in prior conversation-analytic and phonetic work: Competitive, Cooperative, Topic Change, and Backchannel. These intent labels then condition a turn-taking policy over four response strategies: Yield, Resume / Hold, Bridge, and Override (Jeon et al., 7 May 2026).
The paper notes a naming inconsistency across sections but treats “Resume/Hold” as the same operational category. Conceptually, the four actions are distinct. Yield cedes the floor. Resume/Hold continues or completes the prior utterance despite interruption. Bridge acknowledges the interruption while smoothly connecting back or integrating it. Override actively talks over or suppresses the interruption. This makes the action space coarse but interpretable.
The turn-taking manager does not ask the LLM to infer overlap behavior purely from persona text. Instead, the chosen strategy is injected directly into the system prompt as a control token. The paper gives the concrete example:
[STRATEGY=RESUME]: finish your previous sentence, ignoring the user
This mechanism makes turn-taking a pre-committed control variable rather than an emergent side effect of persona prompting. The formal policy itself is probabilistic. In interruption_config.json, researchers specify a matrix mapping interruption intent to a categorical distribution over the four actions. The paper’s example of a dominant persona assigns the following distribution for Competitive interruptions: 50% Resume, 25% Override, 15% Bridge, 10% Yield (Jeon et al., 7 May 2026).
Operationally, the control flow is: the bot is speaking; client VAD detects user speech; client playback halts and the cutoff point is logged; user audio goes to server ASR; zero-shot classification predicts interruption intent; the turn-taking manager consults the persona-conditioned JSON policy; in Style B it samples an action from the categorical distribution; in Style C it does not sample and lets the LLM choose zero-shot from the persona prompt; the selected action is injected into the prompt; the LLM generates recovery or continuation text; and TTS returns the next audio segment (Jeon et al., 7 May 2026).
The paper is explicit about what the framework does not provide. It does not present a formal latency equation, a detailed finite-state machine, explicit timing thresholds, overlap-duration thresholds, hysteresis values, or a learned probabilistic score function. The policy is operational rather than heavily formalized.
4. Session lifecycle, surveys, and A/B testing
PersonaKit is not only a dialogue runtime; it is an end-to-end experimentation framework. A session ends either at a MAX_TURNS cap or when the user expresses a verbal TERMINATE intent. The LLM then emits an in-character farewell containing a hidden [EXIT] tag, which a tag parser intercepts to trigger the post-session survey defined in session_config.json. Exported data include the full dialogue transcript, event logs with per-turn interruption intent and selected strategy, cutoff and remaining text, and all survey responses in JSON or CSV (Jeon et al., 7 May 2026).
The platform automates comparative A/B-style studies after live interaction. The UI supports both desktop and mobile. The participant view displays dialogue turns, persona and style labels, and live VAD status, and the post-session survey is auto-generated from configuration. The survey schema can include arbitrary Likert, forced-choice, and free-text questions, which allows the platform to support constructs beyond the pilot’s main focus.
The paper’s pilot uses three within-subject conditions for each persona: Style A, an Always-Yield baseline; Style B, a probabilistic policy using JSON-tuned strategy weights; and Style C, an autonomous mode where the LLM chooses its own strategy zero-shot from the persona prompt. After each persona interaction, participants complete a comparative Likert survey on the scale {-1, 0, +1} for Reaction Naturalness (“felt human-like and natural”), Persona Consistency (“remained consistent with the role”), and Interaction Fluidity (“turn transitions felt smooth”), along with a forced-choice preference item and a free-text justification (Jeon et al., 7 May 2026).
This survey-centric design is one of the platform’s strongest applied features. It allows a researcher to vary interruption policy while holding the underlying LLM and voice fixed, isolating turn-taking as the manipulated factor.
5. Pilot evaluation and descriptive findings
The pilot evaluation is an in-the-wild, randomized within-subject study over 8 distinct personas, organized by the Interpersonal Circumplex with two personas per quadrant: Q1 High Agency, Low Communion (Drill Sergeant, Tavern Keeper); Q2 High Agency, High Communion (Salesperson, Tour Guide); Q3 Low Agency, High Communion (AI Assistant, Librarian); and Q4 Low Agency, Low Communion (DMV Clerk, Distracted Chef). Five participants completed the full study (N=5), producing 120 dialogue sessions (Jeon et al., 7 May 2026).
The paper explicitly labels the quantitative results as “descriptive, not inferential.” There are no p-values, confidence intervals, standardized effect sizes, ANOVA tables, or significance tests. Means are reported by quadrant with 10 ratings per cell.
The quadrant-level pattern is substantive. For Q1, high-agency, low-communion personas appear to benefit from non-yielding behavior: Reaction Naturalness rises from 0.20 under Yield to 0.60 under Probabilistic, while Autonomous receives 60% of forced-choice preferences. For Q2, the pattern slightly favors the Probabilistic style, which reaches 0.60 naturalness and 50% preference. For Q3, always-yield is the clearest winner: low-agency, highly communal roles such as AI Assistant and Librarian receive 70% preference for Style A, with 0.90 fluidity. For Q4, the pattern is mixed: Yield receives the most preference votes (50%), but Probabilistic attains the highest naturalness (0.67) and consistency (0.90) (Jeon et al., 7 May 2026).
Qualitative comments reinforce the interpretation that interruption behavior contributes directly to persona realism. On the Drill Sergeant, one participant said: “I liked how he ignored me when I interrupted him. Just like a real boot camp sometimes.” On the Tour Guide: “Since it was supposedly an enthusiastic one, B was good because it was trying to finish what it was saying.” On the Tavern Keeper, comments suggested that the autonomous style could feel less impatient or more playful depending on context.
The paper also includes a concrete interruption-recovery trace. In a Drill Sergeant session under the probabilistic condition, the bot intended to say: “Louder, recruit! I can't hear you over your weakness! Repeat it again!” The user interrupted at “Repeat it,” leaving remaining text “again!”. The interruption was classified as COMPETITIVE and sampled as RESUME, so the bot completed with “…again!” This example is used to show how persona-consistent overlap handling can preserve coherence that an always-yield system would discard (Jeon et al., 7 May 2026).
6. Limitations, misconceptions, and significance
PersonaKit’s limitations are explicit and central to its interpretation. The pilot sample is tiny (N=5), so the reported findings are illustrative rather than statistically conclusive. Interruption intent classification is done through a zero-shot LLM prompt and was not validated against human annotations, so classification errors are possible, especially for ambiguous backchannels or noisy audio. The action space is intentionally coarse—Yield, Resume/Hold, Bridge, Override—and does not model finer-grained prosodic or multimodal phenomena such as pitch reset, latching, or gaze. The framework also lacks a formal state machine, explicit timing policy, and inferential statistical analysis (Jeon et al., 7 May 2026).
A common misconception would be to treat PersonaKit as a general computational theory of turn-taking. It is better understood as an experimental research tool. Its contribution is not a new formal model of overlap management; it is the operationalization of turn-taking as a configurable persona variable, together with a deployable workflow for studying that variable under live conditions.
A second misconception would be to regard full-duplex quality as primarily a latency problem. PersonaKit certainly emphasizes low latency through client-side VAD, local playback halt, and byte-level playback tracking, but its deeper argument is sociolinguistic: overlap handling itself shapes perceived realism. This suggests that “always-yield” is not merely suboptimal for some agents; for certain roles it is persona-destroying.
Within the broader persona literature, PersonaKit occupies a distinctive position. PersonaGym evaluates persona-faithfulness across environments and tasks (Samuel et al., 2024); Persona biases work studies harmful differences induced by demographic personas (Sheng et al., 2021); and PersoPilot treats persona as contextualized user modeling for classification and personalization (Afzoon et al., 4 Feb 2026). PersonaKit adds a different research instrument: a low-latency, browser-based full-duplex platform in which interruption policy becomes measurable, comparable, and experimentally manipulable (Jeon et al., 7 May 2026).
The paper’s implied future direction is correspondingly clear: larger-scale studies, better-validated interruption intent models, richer action vocabularies, and learned turn-taking policies that extend beyond coarse symbolic actions into prosodic and multimodal behavior. Within that scope, PersonaKit establishes turn-taking as part of persona design and provides an extensible end-to-end testbed for investigating how conversational floor management shapes the perceived realism of next-generation spoken agents (Jeon et al., 7 May 2026).
