aiqbench analysis
The paper tackles a concrete serving problem: a speech model pairs an autoregressive LLM backbone with a neural audio codec, and existing systems waste GPU capacity by serializing or isolating them. VoxServe splits one GPU into lanes and routes audio chunks by playback deadline, reporting 1.5x the request rate of the best existing system at a 200ms time-to-first-audio target. These results come from one model (Qwen3-TTS), so the gains need independent replication, but the scheduling approach is useful for anyone running streaming speech inference at scale.
VoxServe: Streaming-Centric Serving System for Speech Language Models
arXiv:2602.00269v2 Announce Type: replace-cross Abstract: Speech Language Models (SpeechLMs) are increasingly deployed for streaming speech generation, where audio playback must start quickly and never stall. Serving them requires running two very different model components on the same system. An autoregressive LLM backbone emits audio tokens, and a codec converts chunks of those tokens into waveforms. Existing systems run the two as separately scheduled stages, and therefore either serialize them, let them interfere, or strand capacity by placing them on separate devices. We observe that overlapping the backbone and the codec on separate parts of the GPU raises throughput. Although this increases codec latency, most chunks can tolerate it, as long as it neither lies on the critical path of time-to-first-audio (TTFA) nor causes a playback stall. Based on this insight, we present VoxServe, a SpeechLM serving system that allocates GPU resources according to the urgency of chunks. VoxServe executes work on three asynchronous lanes. The LM lane runs the backbone on a large partition of the GPU's streaming multiprocessors, the background codec lane runs on a small partition beside it, and the urgent lane uses the whole GPU. The scheduler queues each request to the LM lane and routes every chunk the backbone completes by its urgency. A request's first chunk and chunks close to their playback deadline go to the urgent lane, and the rest are batched on the background lane. On Qwen3-TTS, VoxServe sustains 1.5x the request rate of the best existing system at a 200ms TTFA target without playback stalls. The code of VoxServe is available at https://github.com/vox-serve/vox-serve.
Abstract:Speech Language Models (SpeechLMs) are increasingly deployed for streaming speech generation, where audio playback must start quickly and never stall. Serving them requires running two very different model components on the same system. An autoregressive LLM backbone emits audio tokens, and a codec converts chunks of those tokens into waveforms. Existing systems run the two as separately scheduled stages, and therefore either serialize them, let them interfere, or strand capacity by placing them on separate devices. We observe that overlapping the backbone and the codec on separate parts of the GPU raises throughput. Although this increases codec latency, most chunks can tolerate it, as long as it neither lies on the critical path of time-to-first-audio (TTFA) nor causes a playback stall. Based on this insight, we present VoxServe, a SpeechLM serving system that allocates GPU resources according to the urgency of chunks. VoxServe executes work on three asynchronous lanes. The LM lane runs the backbone on a large partition of the GPU's streaming multiprocessors, the background codec lane runs on a small partition beside it, and the urgent lane uses the whole GPU. The scheduler queues each request to the LM lane and routes every chunk the backbone completes by its urgency. A request's first chunk and chunks close to their playback deadline go to the urgent lane, and the rest are batched on the background lane. On Qwen3-TTS, VoxServe sustains 1.5x the request rate of the best existing system at a 200ms TTFA target without playback stalls. The code of VoxServe is available at this https URL.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Distributed, Parallel, and Cluster Computing (cs.DC); Sound (cs.SD); Audio and Speech Processing (eess.AS)
arXiv:2602.00269 [cs.LG]
(or arXiv:2602.00269v2 [cs.LG] for this version)
Full text under the publisher's open licence · 285 words