SpeechGym: An Audio-Native Gym for Training Voice Agents via Reinforcement Learning

• URL: http://arxiv.org/abs/2608.26432v1
• Date: Wed, 26 Aug 2026 22:16:05 GMT
• Title: SpeechGym: An Audio-Native Gym for Training Voice Agents via Reinforcement Learning
• Abstract summary: SpeechGym is an audio-native agentic environment in which two omni-modal models converse in native audio.<n>We present SpeechGym, an audio-native agentic environment in which two omni-modal models converse in native audio.
• Score: 24.07999549415095
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Voice agents must call tools and hold multi-turn dialogue entirely through speech, yet the dominant paradigm trains them in text. Existing frameworks either cascade TTS and ASR around a proprietary voice API, where gradients cannot flow and per-call cost makes on-policy reinforcement learning prohibitive, or stay in text: they measure voice agents but cannot improve them. We present SpeechGym, an audio-native agentic environment in which two omni-modal models converse in native audio, with no external ASR or TTS and no API boundary, over the unmodified tasks, tools and success check of an established text agentic benchmark, so that the interaction modality is the only variable and the loop stays local and trainable end to end. Audio agentic capability does not follow from audio understanding. The failures speech introduces are perceptual rather than reasoning deficits: the agent picks the right tool and the right argument slot but fills it with a value misheard from the waveform, and that single error cascades into a failed call, a retry of the same call, and a wasted step budget. A second failure is behavioural: under an insistent caller the agent performs an unauthorised write and ends the episode believing it helped. Both are trainable, because the environment labels them for free: a call with a misheard argument fails against the database while a correct one succeeds. The obstacle is sparsity, not signal. Outcome-only GRPO is gradient-starved here, since almost every rollout group fails identically, while a per-turn process reward crediting each successful tool call restores variance to nearly every group. Trained this way, the agent transfers with no further tuning to an independently implemented voice benchmark, more than doubling task success and carrying an open-weights model from last place to second on that leaderboard, while using fewer turns and tokens than before training.

Related papers

• AgenticASR: Refining Speech Recognition in Real-World Scenarios via an Agentic Approach [15.617770133323077]
  Agentic Speech Recognition (AgenticASR) is an audio-to-clean-text task.<n>It removes disfluencies, resolves self-corrections, and normalizes written form.<n>AgenticASR attains the highest AASR-Bench scores among evaluated systems.
  arXiv  Detail & Related papers  (2026-07-30T13:12:25Z)
• VAmoS Bench: Voice Agent Simulation Bench [8.26112859764927]
  VAmoS Bench measures complete voice-agent systems end to end in a stateful customer-support task.<n>Each of 100 scenarios supplies a simulated caller with a private goal and a seeded backend.<n>The platform uses each scenario to populate and activate an isolated simulation in which the caller reaches Riley over audio.
  arXiv  Detail & Related papers  (2026-07-29T20:42:38Z)
• UAF: A Unified Audio Front-end LLM for Full-Duplex Speech Interaction [7.775050285048427]
  Speech assistant is driving artificial propagation toward human-like conversational systems.<n>Front-end components, such as voice activity detection (VAD) and turn-taking detection (TD) are crucial for speech assistant.<n>We propose the first unified audio frontend LLM (UAF) tailored for fullgressive speech systems.
  arXiv  Detail & Related papers  (2026-04-21T08:24:55Z)
• Fun-Audio-Chat Technical Report [71.07966678560291]
  temporal resolution between speech tokens (25Hz) and text tokens (3Hz) mitigates semantic information mismatchs and incurs high computational costs.<n>We introduce Fun-Audio-Chat, a Large-stage Spoken Speech-to-scale tasks.<n>Fun-Audio-Chat 8B and MoE 30BA3B achieve competitive performance on SpeechText and Speech-to-scale tasks.
  arXiv  Detail & Related papers  (2025-12-23T08:35:27Z)
• Cloning a Conversational Voice AI Agent from Call\,Recording Datasets for Telesales [0.0]
  We present a methodology for cloning a conversational voice AI agent from a corpus of call recordings.<n>Our system listens to customers over the telephone, responds with a synthetic voice, and follows a structured playbook learned from top performing human agents.<n>The cloned agent is evaluated against human agents on a rubric of 22 criteria covering introduction, product communication, sales drive, objection handling, and closing.
  arXiv  Detail & Related papers  (2025-09-05T07:36:12Z)
• Distilling an End-to-End Voice Assistant Without Instruction Training Data [53.524071162124464]
  Distilled Voice Assistant (DiVA) generalizes to Question Answering, Classification, and Translation. We show that DiVA better meets user preferences, achieving a 72% win rate compared with state-of-the-art models like Qwen 2 Audio.
  arXiv  Detail & Related papers  (2024-10-03T17:04:48Z)
• Multi-modal Adversarial Training for Zero-Shot Voice Cloning [9.823246184635103]
  We propose a Transformer encoder-decoder architecture to conditionally discriminate between real and generated speech features. We introduce our novel adversarial training technique by applying it to a FastSpeech2 acoustic model and training on Libriheavy, a large multi-speaker dataset. Our model achieves improvements over the baseline in terms of speech quality and speaker similarity.
  arXiv  Detail & Related papers  (2024-08-28T16:30:41Z)
• Towards Unsupervised Speech Recognition Without Pronunciation Models [57.222729245842054]
  In this article, we tackle the challenge of developing ASR systems without paired speech and text corpora.<n>We experimentally demonstrate that an unsupervised speech recognizer can emerge from joint speech-to-speech and text-to-text masked token-infilling.<n>This innovative model surpasses the performance of previous unsupervised ASR models under the lexicon-free setting.
  arXiv  Detail & Related papers  (2024-06-12T16:30:58Z)
• REBORN: Reinforcement-Learned Boundary Segmentation with Iterative Training for Unsupervised ASR [54.64158282822995]
  We propose REBORN,Reinforcement-Learned Boundary with Iterative Training for Unsupervised ASR. ReBORN alternates between training a segmentation model that predicts the boundaries of the segmental structures in speech signals and training the phoneme prediction model, whose input is the speech feature segmented by the segmentation model, to predict a phoneme transcription. We conduct extensive experiments and find that under the same setting, REBORN outperforms all prior unsupervised ASR models on LibriSpeech, TIMIT, and five non-English languages in Multilingual LibriSpeech.
  arXiv  Detail & Related papers  (2024-02-06T13:26:19Z)
• Bridging Speech and Textual Pre-trained Models with Unsupervised ASR [70.61449720963235]
  This work proposes a simple yet efficient unsupervised paradigm that connects speech and textual pre-trained models. We show that unsupervised automatic speech recognition (ASR) can improve the representations from speech self-supervised models. Notably, on spoken question answering, we reach the state-of-the-art result over the challenging NMSQA benchmark.
  arXiv  Detail & Related papers  (2022-11-06T04:50:37Z)
• VQMIVC: Vector Quantization and Mutual Information-Based Unsupervised Speech Representation Disentanglement for One-shot Voice Conversion [54.29557210925752]
  One-shot voice conversion can be effectively achieved by speech representation disentanglement. We employ vector quantization (VQ) for content encoding and introduce mutual information (MI) as the correlation metric during training. Experimental results reflect the superiority of the proposed method in learning effective disentangled speech representations.
  arXiv  Detail & Related papers  (2021-06-18T13:50:38Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.