VCB-Bench: An Evaluation Benchmark for Audio-Grounded Large Language Model Conversational Agents
Introduction
Voice Chat Bot Bench (VCB Bench) is a high-quality Chinese benchmark built entirely on real human speech. It evaluates large audio language models (LALMs) along three complementary dimensions:
(1) Instruction following: Text Instruction Following (TIF), Speech Instruction Following (SIF), English Text Instruction Following (TIF-En), English Speech Instruction Following (SIF-En) and Multi-turn Dialog (MTD);
(2) Knowledge: General Knowledge (GK), Mathematical Logic (ML), Discourse Comprehension (DC) and Story Continuation (SC).
(3) Robustness: Speaker Variations (SV), Environmental Variations (EV), and Content Variations (CV).
Getting Started
Installation:
git clone https://github.com/Tencent/VCB-Bench.git
cd VCB-Bench
pip install -r requirements.txt
Note: To evaluate Qwen3-omni, please replace it with the environment it requires.
Download Dataset:
Download the dataset from Hugging Face and place the 'vcb_bench' into 'data/downloaded_datasets'.
Evaluation:
This code is adapted from Kimi-Audio-Evalkit, where you can find more details about the evaluation commands.
(1) Inference + Evaluation:
python run_audio.py --model {model_name} --data {data_name}
For example:
CUDA_VISIBLE_DEVICES=1 python run_audio.py --model Qwen2.5-Omni-7B --data general_knowledge
(2) Only Inference:
python run_audio.py --model {model_name} --data {data_name} --skip-eval
For example:
CUDA_VISIBLE_DEVICES=4,5,6,7 python run_audio.py --model StepAudio --data continuation_en creation_en empathy_en recommendation_en rewriting_en safety_en simulation_en emotional_control_en language_control_en non_verbal_vocalization_en pacing_control_en style_control_en volume_control_en --skip-eval
(3) Only Evaluation:
python run_audio.py --model {model_name} --data {data_name} --reeval
For example:
CUDA_VISIBLE_DEVICES=2 nohup python run_audio.py --model Mimo-Audio --data continuation creation empathy --reeval
(4) Inference + ASR + Evaluation:
python run_audio.py --model {model_name} --data {data_name} --wasr
For example:
CUDA_VISIBLE_DEVICES=3 python run_audio.py --model StepAudio2 --data rewriting safety simulation continuation_en --wasr
Format Result:
python sumup_eval.py --model {model_name}
python sumup_eval.py --model {model_name} --export_excel --output_file my_results.xlsx
Supported Datasets and Models
(1) Locate the dataset you need to evaluate from the Data Name column in the Datasets table, and populate the {data_name} parameter in the evaluation command accordingly.
(2) Each dataset in the SV, EV, and CV sections has a corresponding comparison dataset named "{data_name}_cmp", following the specified naming convention.
(3) Identify the model you intend to evaluate from the Model Name column in the Models table, and insert the appropriate {model_name} into the evaluation command.
Datasets:
| Data Type | Data Name | Detail |
|---|---|---|
| TIF | continuation | - |
| creation | - | |
| empathy | - | |
| recommendation | - | |
| rewriting | - | |
| safety | - | |
| simulation | - | |
| TIF-En | continuation_en | - |
| creation_en | - | |
| empathy_en | - | |
| recommendation_en | - | |
| rewriting_en | - | |
| safety_en | - | |
| simulation_en | - | |
| SIF | emotional_control | - |
| language_control | - | |
| non_verbal_vocalization | - | |
| pacing_control | - | |
| style_control | - | |
| volume_control | - | |
| SIF-En | emotional_control_en | - |
| language_control_en | - | |
| non_verbal_vocalization_en | - | |
| pacing_control_en | - | |
| style_control_en | - | |
| volume_control_en | - | |
| MTD | progression | - |
| backtracking | - | |
| transition | - | |
| GK | general_knowledge | mathematics, geography, politics, chemistry, biology, law, physics, history, medicine, economics, sports, culture |
| ML | basic_math | - |
| math | - | |
| logical_reasoning | analysis, induction, analogy, logic | |
| DC | discourse_comprehension | inference, induction, analysis |
| SV | age | child, elder |
| accent | tianjin, beijing, dongbei, sichuan | |
| volume | down, up | |
| speed | - | |
| EV | non_vocal_noise | echo, outdoors, far_field |
| vocal_noise | TV_playback, background_chat, vocal_music, voice_announcement | |
| unstable_signal | - | |
| CV | casual_talk | - |
| mispronunciation | - | |
| grammatical_error | - | |
| topic_shift | - | |
| code_switching | - |
Models:
| Model Type | Model Name |
|---|---|
| Chat Model | Qwen2-Audio-7B-Instruct |
| Qwen2.5-Omni-7B | |
| Baichuan-Audio-Chat | |
| GLM4-Voice | |
| Kimi-Audio | |
| Mimo-Audio | |
| StepAudio | |
| StepAudio2 | |
| GPT4O-Audio | |
| Qwen3-Omni-Instruct | |
| Pretrain Model | Qwen2-Audio-7B |
| Baichuan-Audio | |
| Kimi-Audio-Base | |
| StepAudio2-Base |
Acknowledge
We borrow some code from Kimi-Audio-Evalkit, GLM-4-Voice, Baichuan-Audio, Kimi-Audio, Mimo-Audio, Step-Audio2, and StepAudio.
Citation
@misc{hu2025vcbbenchevaluationbenchmark,
title={VCB Bench: An Evaluation Benchmark for Audio-Grounded Large Language Model Conversational Agents},
author={Jiliang Hu and Wenfu Wang and Zuchao Li and Chenxing Li and Yiyang Zhao and Hanzhao Li and Liqiang Zhang and Meng Yu and Dong Yu},
year={2025},
eprint={2510.11098},
archivePrefix={arXiv},
primaryClass={cs.SD},
url={https://arxiv.org/abs/2510.11098},
}