The paper introduces a symbiotic architecture that equips large language models with audio‑understanding abilities without fine‑tuning their weights. It uses an injector module to write audio‑conditioned vectors into the LLM’s key‑value cache, allowing the model to act as an audio language model while keeping the backbone unchanged. The approach improves scalability—since injection cost depends on the injector width—and preserves the LLM’s original text performance, outperforming conventional frozen‑LLM methods and approaching fine‑tuned ALM results on audio tasks.
By Yotaro Kubo, Qi Sun, Yujin Tang
arXiv:2609.36577v1 Announce Type: cross
Abstract: Audio large language models (ALLMs) can reason about the content of audio recordings to perform complex tasks. However, these capabilities usually co...
By Zhenhong Zhou, Xuanyue Zhao, Youji Liu, Yuanhe Zhang, Xiaoyu Ma, Lianyu Hu, Yang Liu
arXiv:2608. 08569v1 Announce Type: new Abstract: Recent advancements in Speech Large Language Models have demonstrated remarkable capabilities in understanding complex audio tasks.
By Wenxu Jia, Dongjie Fu, Xize Cheng, Fangming Feng, Linjun Li, Wenshi Chen, Yingming Li, Zhou Zhao, Tao Jin
The paper introduces AWED, a word‑level emission‑delay metric, and demonstrates that post‑training a delayed‑streaming model with a joint reward (GRPO) improves both accuracy and latency. Using a single operating point (τ=6 frames), the method outperforms both its supervised baseline and the Voxtral Realtime backbone across all lookahead budgets, reducing WER by up to 30.8% at 80 ms delay and lowering median AWED from 1.17 s to 1.04 s.
By Xiuwen Zheng
arXiv:2603. 09714v2 Announce Type: replace-cross Abstract: While multi-audio understanding is critical for large audio-language models (LALMs), it remains underexplored.
By Chih-Kai Yang, Yun-Shao Tsai, Yu-Kai Guo, Ping-Le Tsai, Yen-Ting Piao, Hung-Wei Chen, Ting-Lin Hsiao, Yun-Man Hsu, Ke-Han Lu, Hung-yi Lee
arXiv:2606. 05121v1 Announce Type: cross Abstract: Audio is an inherently interactive modality, yet today's Large Audio Language Models (LALMs) are offline, and streaming audio models each handle only a single task such as streaming ASR or voice chatting.
By Zhifei Xie, Zihang Liu, Ze An, Xiaobin Hu, Yue Liao, Ziyang Ma, Dongchao Yang, Mingbao Lin, Deheng Ye, Shuicheng Yan, Chunyan Miao
X-AuT is a progressive framework for compressing the audio encoder of speech large language models. It selects layer combinations via short behavioral probes and restores performance through representation alignment, cross‑scale distillation, scheduled student‑policy supervision, and LoRA finetuning, while keeping the language‑model backbone frozen. On ten Chinese–English benchmarks, reducing Qwen3‑ASR‑0.6B’s encoder from 18 to 16 layers lowers macro‑average error from 5.61% to 5.27%, and a 14‑layer model achieves 5.75% error with 20.7% fewer parameters.
By Haojun Zhang, Yi Zou, Min Chen, Qize Yu, Lianrui Fan, Xini Ding, Hao Li, Shuchang Zhou, Xianming Liu, Shiyu Huang
arXiv:2609.05871v1 Announce Type: cross
Abstract: Audio-conditioned language models often underuse acoustic cues such as prosody, emotion, and non-speech sounds, raising the question of whether ASR-s...
By Song-ha Jo, Sehyun Lee, Soyoon Kim, Jaesik Choi, Sanghyuk Choi
GrainSpeech is a compact speech synthesis model that uses a fixed‑receptive‑field convolutional encoder to reduce pitch, energy, and duration prediction errors by 36.0%, 17.3%, and 3.4% respectively. It introduces a Mel‑specific gradient‑variance supervision that improves fine‑scale variation while avoiding quality degradation. With only 264.8K parameters, GrainSpeech achieves 17.9× real‑time Mel generation on a microcontroller and attains UTMOS scores comparable to much larger models, using less than 1.5% of their parameters.
By Zitao Liang, Chang Gao
TontaubeV1 is a streaming text‑to‑speech model that preserves natural prosody while running on a single consumer GPU. It encodes speech with a hierarchical DualCodec representation at 12.5 Hz, separating a semantic stream from successive acoustic refinements, and uses Qwen3‑derived transformers to predict the semantic stream, utterance duration, and acoustic refinements. The system supports up to one minute of reference audio for voice conditioning, streams with a 200 ms latency to first audio, and achieves real‑time factors of 0.08 (single input) and 0.02 (eight concurrent inputs).
By Fritz Cremer, Jonathan Cremer
TEMPO is a unified model that adds temporally‑grounded capabilities to large audio‑language models, enabling timestamping of events, speakers, and sounds in audio, speech, and music. It introduces a supervised fine‑tuning stage featuring atomic timestamp tokens, a time‑aware projector with sinusoidal encodings, and a distance‑aware Gaussian loss, trained via a synthetic‑to‑real curriculum. Additionally, TEMPO employs reinforcement learning (GRPO) as a refinement step, and achieves state‑of‑the‑art performance on a benchmark of 10K samples across five timestamping tasks, surpassing Audio Flamingo Next and Qwen3‑Omni.
By Apoorva Kulkarni, Kaousheik Jayakumar, Sreyan Ghosh, Utathya Aich, Ramani Duraiswami, Dinesh Manocha
The paper introduces CUES, a lightweight heuristic for selecting encoder combinations in large audio‑language models by estimating complementarity through Pearson correlations of single‑encoder performance profiles. Using a frozen SmolLM2‑135M backbone, CUES consistently identifies optimal encoder sets for each track on the XARES‑LLM benchmark without requiring fusion training or test data. On broad audio tasks, CUES selects a diverse trio of encoders, improving performance by 4.3% over Whisper‑medium, while on text generation it opts for a focused speech‑only pair, outperforming mHuBERT‑147 by 6.3%. The results illustrate how correlation signals guide a diversity–interference trade‑off across different task families.
By Pei-Jun Liao, Hung-Shin Lee, Wenze Ren, Kuo-Hsuan Hung, Hung-yi Lee, Hsin-Min Wang