# Grounded in Sound: Reinforcement Learning with a Frozen Acoustic Judge to Curb ASR Insertion Hallucinations

arXiv: https://arxiv.org/abs/2609.14455 · HTML: https://mlsanity.com/paper/2609.14455/

## Key facts

- No citing papers in the map yet (data through 19 Oct 2026).
- Placed in Speech recognition models and training, one of 11 directions in Speech and audio language models.
- Builds on 23 of its 28 references that are in the map.
- First posted on arXiv 13 Sep 2026.

## Placed in

- Speech and audio language models / [Speech recognition models and training](https://mlsanity.com/speech/speech-recognition-models-and-training/)

## Builds on (most cited first, 23 of 23)

- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [Proximal Policy Optimization Algorithms](https://mlsanity.com/paper/1707.06347/) (2017)
- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Robust Speech Recognition via Large-Scale Weak Supervision](https://mlsanity.com/paper/2212.04356/) (2022)
- [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://mlsanity.com/paper/2006.11477/) (2020)
- [Deep reinforcement learning from human preferences](https://mlsanity.com/paper/1706.03741/) (2017)
- [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://mlsanity.com/paper/2106.07447/) (2021)
- [Selective Classification for Deep Neural Networks](https://mlsanity.com/paper/1705.08500/) (2017)
- [MUSAN: A Music, Speech, and Noise Corpus](https://mlsanity.com/paper/1510.08484/) (2015)
- [Qwen2-Audio Technical Report](https://mlsanity.com/paper/2407.10759/) (2024)
- [SelectiveNet: A Deep Neural Network with an Integrated Reject Option](https://mlsanity.com/paper/1901.09192/) (2019)
- [A Comparison of Techniques for Language Model Integration in Encoder-Decoder Speech Recognition](https://mlsanity.com/paper/1807.10857/) (2018)
- [Whispering LLaMA: A Cross-Modal Generative Error Correction Framework for Speech Recognition](https://mlsanity.com/paper/2310.06434/) (2023)
- [Careless Whisper: Speech-to-Text Hallucination Harms](https://mlsanity.com/paper/2402.08021/) (2024)
- [HyPoradise: An Open Baseline for Generative Speech Recognition with Large Language Models](https://mlsanity.com/paper/2309.15701/) (2023)
- [Hallucinations in Neural Automatic Speech Recognition: Identifying Errors and Hallucinatory Models](https://mlsanity.com/paper/2401.01572/) (2024)
- [Investigation of Whisper ASR Hallucinations Induced by Non-Speech Audio](https://mlsanity.com/paper/2501.11378/) (2025)
- [Lost in Transcription, Found in Distribution Shift: Demystifying Hallucination in Speech Foundation Models](https://mlsanity.com/paper/2502.12414/) (2025)
- [Audio-attention discriminative language model for ASR rescoring](https://mlsanity.com/paper/1912.03363/) (2019)
- [Group Relative Policy Optimization for Speech Recognition](https://mlsanity.com/paper/2509.01939/) (2025)
- [Explore the Reinforcement Learning for the LLM based ASR and TTS system](https://mlsanity.com/paper/2509.18569/) (2025)
- [From Knowing to Doing Precisely: A General Self-Correction and Termination Framework for VLA models](https://mlsanity.com/paper/2602.01811/) (2026)

## Abstract (arXiv)

When reinforcement learning (RL) is used for post-training automatic speech recognition (ASR), the reward almost always lives in the text space: it compares a hypothesis with the reference and never checks whether the hypothesis is supported by the audio. On highly regular speech this licenses a shortcut - guessing from a strong language prior rather than listening. Once the acoustics degrade, the shortcut runs unchecked and emits fluent but ungrounded words, i.e., insertion errors. We propose an acoustic-fidelity reward: a GRPO reward augmented with a separately pretrained, permanently frozen, non-autoregressive character-level wav2vec2-CTC acoustic judge, used strictly at training and absent at inference, where a single model decodes greedily. Trained on LibriSpeech and evaluated across a six-tier difficulty gradient including real AMI meeting speech (33,282 utterance-condition instances), the method reduces insertion errors by 28.3% on close-talking AMI-IHM and 22.3% on far-field AMI-SDM, while lowering WER on AMI-SDM from 35.89% to 34.71% and showing no detectable WER difference on the other five tiers, against a schedule-matched WER-GRPO baseline. The insertion reduction holds under a meeting-level clustered bootstrap. Four prespecified analyses support content-conditioned insertion calibration: output collapses 85-90% on unintelligible audio that preserves energy and voice activity; the gain is not recovered by the evaluated 32-best CTC rescoring configuration, yet RL internalizes it into a single greedy decoding run; and policy-only confidence yields lower insertion-AURC in all four evaluated settings. We frame this as a mechanism paper, demonstrated in one instantiation: a 7B speech LLM with a 0.3B CTC judge.