Reducing the Output-Mode Gap in Speech Language Models via Joint-Output On-Policy Distillation

• URL: http://arxiv.org/abs/2609.15313v1
• Date: Mon, 14 Sep 2026 10:03:49 GMT
• Title: Reducing the Output-Mode Gap in Speech Language Models via Joint-Output On-Policy Distillation
• Abstract summary: Autoregressive generation of interleaved text and acoustic tokens is a common approach to spoken-response generation in speech large language models.<n>We observe markedly lower answer accuracy for the internal text generated in speech-to-text-and-speech mode than for speech-to-text (S2T) responses.<n>To reduce OMG, we propose emphJoint-Output On-Policy Distillation (JO-OPD), which distills the model's stronger S2T policy into joint generation.
• Score: 9.626979634291063
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Autoregressive generation of interleaved text and acoustic tokens is a common approach to spoken-response generation in speech large language models. Although this design enables streaming generation with explicit textual guidance, generated acoustic tokens become part of the context for subsequent text predictions. Given identical speech inputs, we observe markedly lower answer accuracy for the internal text generated in speech-to-text-and-speech (S2TS) mode than for speech-to-text (S2T) responses. We term this discrepancy the \emph{output-mode gap} (OMG). To reduce OMG, we propose \emph{Joint-Output On-Policy Distillation} (JO-OPD), which distills the model's stronger S2T policy into joint generation using student-generated S2TS trajectories. At each text position, the S2T teacher provides soft targets from a text-only projection of the student's preceding outputs, while the student predicts from the corresponding full interleaved history. A preservation objective further regularizes native non-text predictions. Experiments on Step-Audio-2-mini and Baichuan-Audio-Instruct reveal OMG across two interleaved generation architectures. On Step-Audio-2-mini, JO-OPD reduces OMG from 42.87 to 16.26 percentage points on Spoken-MQA and from 29.72 to 13.04 points on speech-rendered GSM8K, with little change in S2T accuracy and substantially larger reductions than matched SFT baselines. ASR-based evaluation further shows a 7.49-point improvement in spoken-answer accuracy on Spoken-MQA.

Related papers

• DiffuSpeech: Silent Thought, Spoken Answer via Unified Speech-Text Diffusion [23.01044837428522]
  We introduce textbfSilent Thought, Spoken Answer'' -- a paradigm where speech LLMs generate internal text reasoning alongside spoken responses.<n>We present method, the first diffusion-based speech-text language model supporting both understanding and generation.<n>Experiments show method achieves state-of-the-art speech-to-speech QA accuracy, outperforming the best baseline by up to 9 points.
  arXiv  Detail & Related papers  (2026-01-30T12:08:33Z)
• IndexTTS2: A Breakthrough in Emotionally Expressive and Duration-Controlled Auto-Regressive Zero-Shot Text-to-Speech [13.3878636941081]
  IndexTTS2 is a novel, general, and autoregressive model-friendly method for speech duration control.<n>It achieves disentanglement between emotional expression and speaker identity, enabling independent control over timbre and emotion.<n>It outperforms state-of-the-art zero-shot TTS models in terms of word error rate, speaker similarity, and emotional fidelity.
  arXiv  Detail & Related papers  (2025-06-23T08:33:40Z)
• GOAT-TTS: Expressive and Realistic Speech Generation via A Dual-Branch LLM [42.93855899824886]
  We propose a text-to-speech generation approach optimized via a novel dual-branch ArchiTecture (GOAT-TTS)<n>GOAT-TTS combines a speech encoder and projector to capture continuous acoustic embeddings, enabling bidirectional correlation between paralinguistic features (language, timbre, emotion) and semantic text representations without transcript dependency.<n> Experimental results demonstrate that our GOAT-TTS achieves performance comparable to state-of-the-art TTS models.
  arXiv  Detail & Related papers  (2025-04-15T01:44:56Z)
• Generative Pre-trained Speech Language Model with Efficient Hierarchical Transformer [39.31849739010572]
  We introduce textbfGenerative textbfPre-trained textbfSpeech textbfTransformer (GPST) GPST is a hierarchical transformer designed for efficient speech language modeling.
  arXiv  Detail & Related papers  (2024-06-03T04:16:30Z)
• Learning Speech Representation From Contrastive Token-Acoustic Pretraining [57.08426714676043]
  We propose "Contrastive Token-Acoustic Pretraining (CTAP)", which uses two encoders to bring phoneme and speech into a joint multimodal space. The proposed CTAP model is trained on 210k speech and phoneme pairs, achieving minimally-supervised TTS, VC, and ASR.
  arXiv  Detail & Related papers  (2023-09-01T12:35:43Z)
• ContextSpeech: Expressive and Efficient Text-to-Speech for Paragraph Reading [65.88161811719353]
  This work develops a lightweight yet effective Text-to-Speech system, ContextSpeech. We first design a memory-cached recurrence mechanism to incorporate global text and speech context into sentence encoding. We construct hierarchically-structured textual semantics to broaden the scope for global context enhancement. Experiments show that ContextSpeech significantly improves the voice quality and prosody in paragraph reading with competitive model efficiency.
  arXiv  Detail & Related papers  (2023-07-03T06:55:03Z)
• SpeechLM: Enhanced Speech Pre-Training with Unpaired Textual Data [100.46303484627045]
  We propose a cross-modal Speech and Language Model (SpeechLM) to align speech and text pre-training with a pre-defined unified representation. Specifically, we introduce two alternative discrete tokenizers to bridge the speech and text modalities. We evaluate SpeechLM on various spoken language processing tasks including speech recognition, speech translation, and universal representation evaluation framework SUPERB.
  arXiv  Detail & Related papers  (2022-09-30T09:12:10Z)
• TranSpeech: Speech-to-Speech Translation With Bilateral Perturbation [61.564874831498145]
  TranSpeech is a speech-to-speech translation model with bilateral perturbation. We establish a non-autoregressive S2ST technique, which repeatedly masks and predicts unit choices. TranSpeech shows a significant improvement in inference latency, enabling speedup up to 21.4x than autoregressive technique.
  arXiv  Detail & Related papers  (2022-05-25T06:34:14Z)
• JETS: Jointly Training FastSpeech2 and HiFi-GAN for End to End Text to Speech [7.476901945542385]
  We present end-to-end text-to-speech (E2E-TTS) model which has a simplified training pipeline and outperforms a cascade of separately learned models. Our proposed model is jointly trained FastSpeech2 and HiFi-GAN with an alignment module. Experiments on LJSpeech corpus shows that the proposed model outperforms publicly available, state-of-the-art implementations of ESPNet2-TTS.
  arXiv  Detail & Related papers  (2022-03-31T07:25:11Z)
• Wav2vec-Switch: Contrastive Learning from Original-noisy Speech Pairs for Robust Speech Recognition [52.71604809100364]
  We propose wav2vec-Switch, a method to encode noise robustness into contextualized representations of speech. Specifically, we feed original-noisy speech pairs simultaneously into the wav2vec 2.0 network. In addition to the existing contrastive learning task, we switch the quantized representations of the original and noisy speech as additional prediction targets.
  arXiv  Detail & Related papers  (2021-10-11T00:08:48Z)
• AdaSpeech 2: Adaptive Text to Speech with Untranscribed Data [115.38309338462588]
  We develop AdaSpeech 2, an adaptive TTS system that only leverages untranscribed speech data for adaptation. Specifically, we introduce a mel-spectrogram encoder to a well-trained TTS model to conduct speech reconstruction. In adaptation, we use untranscribed speech data for speech reconstruction and only fine-tune the TTS decoder.
  arXiv  Detail & Related papers  (2021-04-20T01:53:30Z)
• FastSpeech 2: Fast and High-Quality End-to-End Text to Speech [189.05831125931053]
  Non-autoregressive text to speech (TTS) models such as FastSpeech can synthesize speech significantly faster than previous autoregressive models with comparable quality. FastSpeech has several disadvantages: 1) the teacher-student distillation pipeline is complicated and time-consuming, 2) the duration extracted from the teacher model is not accurate enough, and the target mel-spectrograms distilled from teacher model suffer from information loss. We propose FastSpeech 2, which directly training the model with ground-truth target instead of the simplified output from teacher, and 2) introducing more variation information of speech (e.g., pitch,
  arXiv  Detail & Related papers  (2020-06-08T13:05:40Z)

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.