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
TontaubeV1 is a streaming text‑to‑speech model that maintains 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 and add refinements. The model 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).
GEPARD is a streaming text‑to‑speech model that uses a standard large language model backbone to generate speech autoregressively, decoding audio with an FSQ‑based neural codec. It streams audio chunk‑by‑chunk as text arrives, achieving a real‑time factor of about 0.067 and an aggregate speedup of roughly 204× on a single GPU with 256 concurrent streams. The design keeps all complex auxiliary mechanisms outside the decode loop, enabling deployment with a standard LLM engine (vLLM) without kernel modifications.
By Denis Pavlov, Ulanbek Abdurazakov, Nursultan Bakashov
Mizar is a 159.3‑million‑parameter audio‑language model designed for devices with limited memory and computation. It couples a compact CED‑Small audio encoder with SmolLM2‑135M via a frequency‑merging mapper and is trained in three stages—audio‑language alignment, audio‑dependent fine‑tuning, and post‑training—to improve performance on audio‑question tasks. Across five random seeds, Mizar outperforms all other sub‑200M‑parameter ALMs on MMAU, MMAR, and ADQA‑clean, achieving mean accuracies of 52.92%, 42.42%, and 36.02% respectively, while enabling local inference on a single CPU with an average latency of 1.09 seconds for MMAU questions.
By Kaiyang Li, Shaobo Han, Yue Tian, Shihao Ji
The paper introduces Triage, a method that predicts the attention distribution of audio tokens before a language model processes them, enabling early pruning of less important tokens. By fitting a linear map to encoder outputs, Triage achieves high correlation (ρ ≥ 0.69) with full-model attention across eleven of thirteen large audio language models. Using this prediction, Triage compresses audio inputs while maintaining near‑full performance, outperforming baselines in transcription accuracy and significantly increasing the amount of audio that fits within a model’s context window.
By Kyoungjun Park, Yunzhe Li, Lili Qiu
SpeechGym is an audio‑native environment that lets two omni‑modal models converse entirely in native audio, eliminating external ASR/TTS and API boundaries while preserving the tasks, tools, and success checks of a standard text‑based agent benchmark. By training end‑to‑end, the framework addresses perceptual failures—such as misheard arguments that cascade into failed calls—and behavioural failures, both of which are automatically labeled for free. Using per‑turn process rewards to overcome reward sparsity, agents trained in SpeechGym transfer to an independent voice benchmark, doubling task success and improving efficiency in turns and tokens.
By Jiajun Fan, Jingyuan Li, Prashanth Gurunath Shivakumar, Jia-Hong Huang, Qi Luo, M. Maruf, Ivan Bulyko, Ge Liu, Roger Ren
FireRedAudio is a 9‑billion‑parameter audio language model that separates continuous input representations for audio understanding and speech generation, enabling a single autoregressive LLM to perform tasks such as ASR, zero‑shot TTS, Instruct TTS, and semantic/acoustic speech editing. The model uses a dedicated Audio Encoder for recognition and a RedAE‑based pathway for generation, with the LLM directly generating text or conditioning a flow‑matching DiT to produce acoustic latents. Evaluations show competitive or leading performance in multilingual ASR, content‑accurate zero‑shot TTS, strong instruction following, and significant improvements in speech editing over prior work.
By Feiyu Shen, Fenglong Xie, Junjie Li, Kun Xie, Lei Xie, Xu Tang, Xuelong Geng, Yan Jia, Yao Hu, Yichen Han, Yichen Wu, Ziqi Dai, Junjie Chen, Kai Huang, Manzhen Wei, Yixuan Li
Recent advances in zero-shot text-to-speech (TTS) have substantially improved speech quality and voice cloning fidelity. However, many zero-shot TTS systems still depend on audio prompt transcripts at inference time.
arXiv:2606. 10368v1 Announce Type: cross Abstract: Speech-to-text (S2T) systems for recognition (ASR) and translation (S2TT) typically generate discrete text tokens.
By Xuanchen Li, Tianrui Wang, Yuheng Lu, Zikang Huang, Yu Jiang, Chenghan Lin, Chenrui Cui, Ziyang Ma, Xingyu Ma, Chunyu Qiang, Guochen Yu, Xie Chen, Longbiao Wang, Jianwu Dang
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
SEA-SpeechBench is a large‑scale multitask benchmark for speech understanding in 11 Southeast Asian languages, comprising 97,194 samples across 99 evaluation sets and 597 hours of curated audio. It covers nine tasks in three categories—speech processing, paralinguistic analysis, and a novel temporal understanding dimension—using multilingual prompting in both native SEA languages and English. Evaluation of current models shows significant performance gaps, especially in temporal understanding, emotion recognition, and speech translation, with low‑resource languages lagging behind English by up to 41 percentage points.
By Jingyi Liao, Wenyu Zhang, Zhuohan Liu, Yingxu He, Geyu Lin, Xunlong Zou, Shuo Sun, Syed Ali Redha Alsagoff, Ai Ti Aw
arXiv:2606. 07547v1 Announce Type: cross Abstract: Speech-based large language models are typically constrained to spoken replies, which limits their user-facing outputs to what can be verbalized and suppresses text-native capabilities such as code generation, structured analysis, and multi-step reasoning in realtime interaction, for tasks that require persistent, structured, and inspectable intermediate outputs.
By Luoyuan Zhang, Bokai Xu, Junbo Cui, Weiyue Sun, Yingjing Xu, Hanyu Liu, Yuan Yao