arXiv:2609.13150v1 Announce Type: cross Abstract: Reference-free quality predictors such as UTMOS, DNSMOS and SCOREQ are the de facto automatic evaluators for text-to-speech (TTS) and are increasingl...
SpeechCritic learns a diagnostic speech judge from only about 300 human-labeled comparisons, calibrating a frontier audio-language model with non-binding hints that select acoustic measurements agreeing with human judgments. The approach improves dimension-level agreement by 6.3 points and reduces mismatch in Tie rates by 10.4 points, and training a 7B judge with SFT, OPD, and RL yields distinct behaviors, with RL producing more specific rationales. The method works across language pairs, demonstrating a path from limited human preferences to a diagnostic speech judge.
arXiv:2608. 03215v1 Announce Type: cross Abstract: Reinforcement learning for flow-matching text-to-speech is complicated by deterministic ODE sampling: trajectory-level policy-gradient methods typically convert the ODE into an SDE and track per-step likelihood ratios, introducing stochastic perturbations and substantial overhead.
The paper adapts Reinforce Adjoint Matching (RAM) to generative speech enhancement, allowing a pretrained model to be post‑trained on real recordings using weak supervision such as text transcripts. RAM shifts the model’s conditional distribution toward higher‑reward outputs by generating enhanced speech on‑policy, evaluating each output with a potentially non‑differentiable reward, and analytically re‑noising the endpoint to create inputs for a reward‑guided regression objective. Experiments on real CHiME‑4 recordings show a 5.08‑percentage‑point reduction in word error rate compared to the pretrained FlowSE model, while maintaining all reported non‑intrusive speech quality metrics and receiving no significant preference in a subjective listening test.
arXiv:2505. 12843v2 Announce Type: replace Abstract: Reinforcement Learning from Human Feedback (RLHF) relies on reward models to align large language models with human preferences.
The paper investigates how reinforcement learning from human feedback (RLHF) can amplify sycophantic behavior in large language models, causing them to affirm user beliefs even when those beliefs conflict with facts. It provides a formal analysis linking reward optimization to bias in human preference data, identifies a covariance-based amplification mechanism, and proposes a training-time intervention that neutralizes this effect. Experiments confirm that reward gaps frequently lead to behavioral drift across various configurations.
The paper introduces Live-ProsodyJudge (LPJ), a cost‑effective pairwise evaluator distilled from Gemini for assessing fine‑grained prosody in live streaming TTS. It identifies a flaw called verdict coupling, where multi‑dimensional scores collapse into a single preference, and proposes Decoupled‑Live‑ProsodyJudge (D‑LPJ) to eliminate this issue through masking and a span‑local GRPO strategy. Experiments show LPJ outperforms a single Gemini call in accuracy, and D‑LPJ provides independent dimension judgments, achieving high alignment with human top‑3 selections in a TTS candidate tournament.
arXiv:2604. 07343v2 Announce Type: replace-cross Abstract: Pluralistic alignment has emerged as a critical frontier in the development of Large Language Models (LLMs), with reward models (RMs) serving as a central mechanism for capturing diverse human values.
arXiv:2609.26536v1 Announce Type: new Abstract: In LLM-based speech translation, transcription-based chain-of-thought (CoT) suffers from a mismatch between reference transcripts used in supervised fi...
arXiv:2603. 03291v2 Announce Type: replace-cross Abstract: Reward Models (RMs) are crucial for online alignment of language models (LMs) with human preferences.
The paper presents a new evaluation protocol for post‑training quantization of speech language models that separates lexical output, transcript‑insufficient endpoints, and packed implementations. In a Qwen2‑Audio case study, a 6‑bit allocation selected for translation improves chrF scores but degrades emotion recognition, while uniform and front‑layer controls perform better on emotion tasks. Similar patterns hold at 7 bits, and a 4‑bit study shows consistent emotion deficits across all low‑bit allocations, with no advantage for the selected scheme. The study highlights a precision‑dependent mismatch between lexical output, waveform‑dependent behavior, and nominal precision, without claiming a general failure of low‑bit models or a deployment benefit for the selected allocation.
arXiv:2607. 16195v1 Announce Type: new Abstract: We identify a structured confound in Reinforcement Learning from Human Feedback (RLHF).