BlogLLM Benchmark Scores as Published, Side by Side
VoiceBench: Results as Published
Every result the catalog holds for VoiceBench: 2 models from 2 developers, read from 1 page. Each number is as printed by the company named beside it; nothing is ranked or rescored. Models are listed by developer, newest first.
Area Multimodal UnderstandingMaintainer Yiming Chen, Xianghu Yue, Chen Zhang, Xiaoxue Gao, Robby T. Tan, and Haizhou LiIntroduced 2024Data version 2026103001, as of 2026-10-30
VoiceBench evaluates voice assistants on spoken instructions rather than on clean text alone. It mixes real and synthetic speech and varies speaker, environment, and content, covering general knowledge, instruction following, and safety. Open-ended answers get a 1 to 5 judge score, reference and multiple-choice sets use accuracy, instruction following averages strict and loose accuracy, and the safety set uses a refusal rate. The paper reports those subset scores for both text-form and speech-form instructions.
VoiceBench
| Model | Value | Printed by | Setup as printed |
|---|---|---|---|
| Gemini 3.1 Pro PreviewGoogle, released 2026-02-19 | 94.3% | Thinking Machines, Thinking Machines Inkling model cardchecked 2026-10-15 | Effort: high |
| InklingThinking Machines, released 2026-07-14 | 91.4% | Thinking Machines, Thinking Machines Inkling model cardchecked 2026-10-15 | Effort: 0.99 |
Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.