DiscoDrive: Synthetic In-Car Dialogue Corpus
- DiscoDrive is a synthetic multi-turn dialogue corpus that models spontaneous driver disfluencies, such as hesitations and false starts, in realistic automotive settings.
- The corpus comprises 3,500 dialogs across seven domains, generated through a two-step DRIVE framework that integrates disfluencies during dialogue simulation.
- Evaluation shows DiscoDrive enhances naturalness and coherence in in-car conversational AI, outperforming traditional fluent datasets like KVRET.
DiscoDrive is a synthetic multi-turn dialogue corpus for in-car conversational AI, introduced together with the DRIVE generation framework in “DRIVE: Disfluency-Rich Synthetic Dialog Data Generation Framework for Intelligent Vehicle Environments” (Chavda et al., 26 Jul 2025). It is designed to model spontaneous disfluencies in driver–assistant interaction, including hesitations, filler words, repetitions, false starts, pauses, and self-corrections. In the paper’s terminology, DRIVE denotes the two-stage prompt-driven generation framework, whereas DiscoDrive denotes the resulting corpus. The resource is positioned as a response to a specific deficiency in automotive conversational datasets: existing corpora are described as too fluent and structurally clean relative to real driver speech, particularly under multitasking, urgency, and cognitive load.
1. Definition and research context
DiscoDrive stands for DIsfluency-Enriched Synthetic COnversations for DRIVEr-Car AI (Chavda et al., 26 Jul 2025). Its immediate research setting is the in-car conversational assistant, where the system must support multi-turn spoken interaction for navigation, weather, vehicle diagnostics, entertainment, car controls, and safety-related requests. The paper argues that this setting is unusually disfluency-prone because drivers speak spontaneously while managing other tasks, and because requests are often revised online rather than composed in a fluent textual form.
A central distinction in the work is between dataset and generation method. DiscoDrive is the corpus; DRIVE is the framework used to generate it. A related misconception addressed implicitly by the paper is that disfluency augmentation can be treated as a purely post-hoc textual corruption problem. The authors instead argue that context-appropriate driver disfluency should be integrated during dialogue realization, not inserted after fluent utterances have already been produced. This methodological claim structures the entire corpus design (Chavda et al., 26 Jul 2025).
The corpus is also positioned against two reference points. First, KVRET is treated as the main automotive baseline, but one that is “largely fluent” and narrower in domain scope. Second, post-hoc disfluency methods such as LARD are presented as less contextually appropriate than dynamic integration during generation. The claimed gap is therefore not merely the absence of synthetic automotive dialogue, but the absence of automotive, multi-domain, disfluency-rich dialogue generated with disfluencies integrated into the generation process itself (Chavda et al., 26 Jul 2025).
2. Corpus composition and task scope
DiscoDrive contains 3,500 dialogs across seven automotive domains, generated from 500 scenarios per domain, with dialog lengths varying across 6, 8, 10, 12, and 14 turns and averaging 8 turns per dialog (Chavda et al., 26 Jul 2025). The paper does not report a formal train/dev/test split for DiscoDrive, and it does not provide a formal machine-readable annotation schema, dialogue-act inventory, or slot ontology.
| Domain | Scope |
|---|---|
| Navigation | Routes, traffic, nearby places |
| Car Maintenance and Diagnostics | Vehicle status, servicing, troubleshooting |
| Safety and Emergency Assistance | Alerts, urgent support, nearby help |
| Entertainment | Media controls, music, radio, audiobooks |
| Local and On-Route Attractions and Activities | Restaurants, attractions, events |
| Car Functions | Air conditioning, seat settings, cabin features |
| Weather | Current weather, route forecasts, alerts |
The role structure alternates between Driver and Car AI. The driver turns are intentionally informal and disfluent; the Car AI turns are shorter, clearer, and task-focused. Disfluencies are represented directly in the textual surface form of utterances through items such as “um,” “uh,” repetitions, ellipses, abrupt restarts, and explicit correction markers such as “no, wait” or “I mean.” The paper does not report a separate token-level disfluency annotation layer (Chavda et al., 26 Jul 2025).
Several example utterances illustrate the intended register. In safety and emergency assistance, the driver says: “Um, can you ... show me a list of nearby charging stations, please?” In weather: “Um, how’s the, uh, weather forecast looking for our route from Surat to Daman, specifically around the coastal areas?” In car maintenance and diagnostics: “Wait, that’s ... um, a little soon, right?” These examples indicate that the corpus emphasizes disfluency primarily on the driver side rather than symmetrically across both interlocutors (Chavda et al., 26 Jul 2025).
3. DRIVE generation framework
The DRIVE framework is a two-step prompt-driven synthetic dialogue generation pipeline consisting of Conversation Scenario Generation followed by Dialog Simulation (Chavda et al., 26 Jul 2025). In the first stage, GPT-4o generates domain-grounded conversation scenarios. For each domain, the scenario generator is prompted with 10–20 human-written examples per domain in a few-shot configuration to encourage contextual relevance, diversity, and domain alignment.
In the second stage, Llama-3.1-8B-Instruct simulates the multi-turn conversation. Generation alternates between the Driver and Car AI roles, with prompts tailored to each speaker and conditioned on recent history. The paper states that conversation history is limited to the last six exchanges in order to preserve coherence while managing computational efficiency. The scenario therefore functions as a semantic premise, while the rolling local history functions as the immediate discourse state (Chavda et al., 26 Jul 2025).
The driver prompt explicitly instructs the model to ask a short, relevant follow-up question and to incorporate disfluency dynamically. The prompt allows one disfluency as a default expectation, while permitting none or multiple disfluencies when this better matches the flow. It also requires that the disfluency remain natural rather than making the utterance incomprehensible. By contrast, the Car AI prompt is constrained to provide accurate, clear, task-specific, concise, and friendly assistance, with minimal unnecessary detail (Chavda et al., 26 Jul 2025).
For technically oriented readers, an important negative fact is that the paper does not provide algorithmic pseudocode, formal objectives, or mathematical equations for the DRIVE generation process. Control is implemented through prompt engineering, role-specific conditioning, scenario grounding, and local conversational context, rather than through an explicit symbolic planner or a formal latent dialogue-state model (Chavda et al., 26 Jul 2025).
4. Disfluency model and methodological distinction
The disfluency taxonomy used in the paper contains five categories: Repetitions, False Starts, Filler Words, Pauses, and Corrections (Chavda et al., 26 Jul 2025). These are defined with short examples such as “I think, I think we should take the next exit” for repetition, “We could-actually, let’s try the other route” for false starts, and “Turn left-no, wait, I mean right” for corrections.
The paper’s strongest methodological claim is that disfluencies are integrated during dialogue generation rather than inserted afterward. This distinction matters because post-hoc insertion is said to risk damaging local coherence, pragmatic appropriateness, timing, and semantic fit. In DRIVE, the utterance is generated as disfluent from the outset, so hesitation or repair is part of the utterance plan rather than an exogenous edit. This suggests a commitment to modeling disfluency as a discourse-level realization phenomenon rather than as superficial surface corruption (Chavda et al., 26 Jul 2025).
The comparison with LARD makes this distinction concrete. LARD is described as applying random token-level or phrase-level operations after fluent dialogue generation, often without full conversational context. DiscoDrive’s alternative is dynamic disfluency integration guided by the scenario, conversation history, and speaker role. The paper’s later human evaluation reports that this dynamic approach is perceived as more natural and more context-appropriate, while clarity remains roughly comparable (Chavda et al., 26 Jul 2025).
The resulting corpus therefore occupies a specific methodological niche. It is neither a purely fluent automotive dataset nor a fluent corpus retrofitted with disfluency markers; rather, it is a synthetic conversational resource in which driver disfluency is a first-class generation constraint.
5. Evaluation protocol and downstream utility
The downstream task used to evaluate DiscoDrive is car AI response generation. Experiments use DialoGPT-Medium and T5-Base, with testing performed on curated in-car-relevant subsets from MultiWOZ 2.2 and Schema-Guided Dialogue (SGD), each containing 220 dialogs (Chavda et al., 26 Jul 2025). Full-data experiments compare zero-shot performance, fine-tuning on the full KVRET training split (2,424 dialogs), and fine-tuning on the full DiscoDrive corpus (3,500 dialogs). Low-resource experiments compare 10% KVRET only (242 dialogs) against 10% KVRET + full DiscoDrive.
The reported automatic metrics are BLEU-1, BLEU-2, BLEU-3, BLEU-4, ROUGE-L, METEOR, BERTScore F1, and N-Distinct (Chavda et al., 26 Jul 2025). On lexical diversity, DiscoDrive exceeds KVRET at all reported n-gram levels: 0.0124 vs 0.0109 for 1-gram, 0.1234 vs 0.1040 for 2-gram, 0.3428 vs 0.2902 for 3-gram, and 0.5425 vs 0.4808 for 4-gram.
In full-data fine-tuning on the MultiWOZ 2.2 in-car subset, DialoGPT-Medium improves from BLEU-4 0.03 and BERTScore F1 82.71 under KVRET fine-tuning to BLEU-4 0.64 and BERTScore F1 85.20 under DiscoDrive fine-tuning. For T5-Base, KVRET fine-tuning yields BLEU-4 0.61, ROUGE-L 13.05, METEOR 9.28, and BERTScore F1 84.13, whereas DiscoDrive fine-tuning yields BLEU-4 0.77, ROUGE-L 16.19, METEOR 22.78, and BERTScore F1 85.48. On the SGD subset, DialoGPT-Medium improves from BLEU-4 0.02 and BERTScore F1 81.23 to BLEU-4 0.28 and BERTScore F1 84.71, while T5-Base improves from BLEU-4 0.67 and BERTScore F1 83.83 to BLEU-4 0.90 and BERTScore F1 85.55 (Chavda et al., 26 Jul 2025).
In low-resource augmentation, gains are often substantial. On the SGD subset, DialoGPT-Medium moves from BLEU-4 0.00, ROUGE-L 5.83, METEOR 4.44, and BERTScore F1 80.44 under 10% KVRET only to BLEU-4 0.88, ROUGE-L 14.02, METEOR 13.46, and BERTScore F1 85.37 with 10% KVRET + DiscoDrive. For T5-Base, the corresponding configuration improves from BLEU-4 1.06, ROUGE-L 7.80, METEOR 5.46, and BERTScore F1 83.36 to BLEU-4 1.26, ROUGE-L 16.60, METEOR 20.40, and BERTScore F1 85.83 (Chavda et al., 26 Jul 2025).
The paper also notes an important nuance: synthetic augmentation does not improve every metric in every model-setting pair. On the MultiWOZ 2.2 subset for T5-Base, BLEU-4 is reported as 1.97 for 10% KVRET only and 1.04 for 10% KVRET + Synthetic, even though BLEU-2, BLEU-3, ROUGE-L, METEOR, and BERTScore F1 all improve substantially. The paper therefore supports a strong but not universal augmentation claim: DiscoDrive improves most reported metrics in most evaluated settings, but not every individual metric monotonically (Chavda et al., 26 Jul 2025).
The experimental infrastructure is also reported. Generation and evaluation use NVIDIA A40 and A100 GPUs, 48 GB and 80 GB VRAM, and bfloat16 precision. The paper reports 40 GPU-hours for scenario and dialog synthesis and 30 GPU-hours for downstream fine-tuning and inference, implemented with Python, HuggingFace Transformers v4.x, and PyTorch (Chavda et al., 26 Jul 2025).
6. Human evaluation, limitations, and significance
Human evaluation is conducted by two human evaluators on stratified subsets of 140 dialogues from DiscoDrive and 140 dialogues from KVRET, using a 5-point Likert scale for intrinsic and disfluency-focused evaluations (Chavda et al., 26 Jul 2025). On intrinsic quality, DiscoDrive scores higher on Naturalness (3.8 ± 0.18 vs 3.6 ± 0.18) and Coherence (4.1 ± 0.16 vs 4.0 ± 0.17), while KVRET scores slightly higher on Engagement (4.0 ± 0.17 vs 3.8 ± 0.17), Consistency (4.3 ± 0.15 vs 4.2 ± 0.16), and On-topic (4.9 ± 0.04 vs 4.7 ± 0.10). The paper interprets this as a realism-versus-structure trade-off.
Pairwise comparative evaluation strengthens the claim that DiscoDrive improves driver-side realism. It defeats KVRET on naturalness (118 vs. 22), human-likeness of driver utterance (123 vs. 17), and engagement (91 vs. 49), while KVRET retains a slight edge on overall quality (72 vs. 68). On task effectiveness, DiscoDrive remains competitive (73 vs. 67) (Chavda et al., 26 Jul 2025).
The direct comparison with KVRET + LARD is especially important for understanding the paper’s methodological thesis. DiscoDrive scores 4.2 ± 0.16 vs 3.6 ± 0.17 on Naturalness and 4.3 ± 0.15 vs 3.4 ± 0.18 on Appropriateness, whereas Clarity is 4.0 ± 0.16 for DiscoDrive and 4.2 ± 0.16 for LARD. The result is not that dynamic integration dominates on every criterion, but that it improves contextual fit and perceived naturalness without a large collapse in clarity (Chavda et al., 26 Jul 2025).
Several limitations are explicit. DiscoDrive is English-only. The authors note that disfluencies may reduce clarity in critical domains such as navigation, safety, and emergencies. The resource remains synthetic even when dynamically generated, and may therefore fail to capture the full unpredictability of real driver speech. The generation pipeline depends on GPT-4o and Llama-3.1-8B-Instruct, making it comparatively resource-intensive. The paper also does not report inter-annotator agreement statistics such as Cohen’s kappa or Krippendorff’s alpha (Chavda et al., 26 Jul 2025).
Within those limits, DiscoDrive functions as a training corpus, an augmentation resource, and a robustness-testing benchmark for in-car conversational systems. Its most specific technical contribution is to operationalize the idea that driver disfluency should be modeled as part of utterance generation itself. That choice, rather than the mere presence of synthetic automotive dialogue, is what distinguishes DiscoDrive within the emerging literature on conversational AI for intelligent vehicle environments (Chavda et al., 26 Jul 2025).