Real-time spoken dialogue systems face a fundamental tension between latency and response quality. End-to-end speech-to-speech (S2S) models respond immediately and naturally handle turn-taking, backchanneling, and interruption, but produce semantically weaker outputs. Cascaded pipelines (ASR ➔ LLM) deliver stronger responses at the cost of latency that grows with model size. We present RelayS2S, a hybrid architecture that runs two paths in parallel upon turn detection. The fast path—a duplex S2S model—speculatively drafts a short response prefix that is streamed immediately to TTS for low-latency response onset, while continuing to monitor live audio events. The slow path—a cascaded ASR ➔ LLM pipeline—generates a higher-quality continuation conditioned on the committed prefix, producing an uninterrupted utterance. A lightweight learned verifier gates the handoff, committing the prefix when appropriate or falling back gracefully to the cascaded pipeline. With GPT-4.1 as the back-end, RelayS2S substantially reduces response latency while preserving nearly all of the cascaded pipeline’s textual quality. On synthetic voice dialogues, it achieves a P90 first-chunk latency of 81 ms, excluding TTS and network latency, compared with 1,006 ms for the cascaded baseline. On real voice dialogues, RelayS2S reduces average first-chunk latency by 479 ms while retaining 99% of the cascaded pipeline’s textual quality. These benefits become larger as the slow-path model scales. Because the prefix handoff requires no architectural modification to either component, RelayS2S serves as a lightweight, drop-in addition to existing cascaded pipelines. Our code is publicly available at https://github.com/mailong25/relays2s.

Neural text generation models have achieved remarkable success in carrying on short open-domain conversations. However, their performance degrades significantly in the long term, especially in their ability to ask coherent questions. A significant issue is the generation of redundant questions where the answer has already been provided by the user. We adapt and evaluate different methods, including negative training, decoding, and classification, to mitigate the redundancy problem. We also propose a simple yet effective method for generating training data without the need for crowdsourcing human-human or human-bot conversations. Experiments with the BlenderBot model show that our combined method significantly reduces the rate of redundant questions from 27.2% to 8.7%, while improving the quality of the original model. The code, dataset, and trained models can be found at our repository.

A principle behind dozens of attribution methods is to take the prediction difference between before-and-after an input feature (here, a token) is removed as its attribution. A popular Input Marginalization (IM) method (Kim et al., 2020) uses BERT to replace a token, yielding more plausible counterfactuals. While Kim et al., 2020 reported that IM is effective, we find this conclusion not convincing as the Deletion-BERT metric used in their paper is biased towards IM. Importantly, this bias exists in Deletion-based metrics, including Insertion, Sufficiency, and Comprehensiveness. Furthermore, our rigorous evaluation using 6 metrics and 3 datasets finds no evidence that IM is better than a Leave-One-Out (LOO) baseline. We find two reasons why IM is not better than LOO: (1) deleting a single word from the input only marginally reduces a classifier’s accuracy; and (2) a highly predictable word is always given near-zero attribution, regardless of its true importance to the classifier. In contrast, making LIME samples more natural via BERT consistently improves LIME accuracy under several ROAR metrics.