| | --- |
| | license: apache-2.0 |
| | datasets: |
| | - KE-Team/SemanticVAD |
| | language: |
| | - zh |
| | - en |
| | base_model: |
| | - Qwen/Qwen2.5-0.5B-Instruct |
| | --- |
| | |
| | # 模型简介 🚀 |
| | 本模型是基于Qwen2.5-0.5B-Instruct架构微调的语义级语音活动检测(Semantic Voice Activity Detection)模型，用于支持全双工语音对话系统。 |
| |
|
| | # 测试集表现 📈 |
| | | 标签 | 准确率 / % | |
| | | :----- | :---------- | |
| | | <打断> | 98.07 | |
| | | <附和> | 98.12 | |
| | | <完成> | 92.73 | |
| | | <未完> | 99.91 | |
| |
|
| |
|
| | # 基础使用 🛠️ |
| |
|
| | ```python |
| | from transformers import AutoTokenizer, AutoModelForCausalLM |
| | |
| | # 加载模型 |
| | model = AutoModelForCausalLM.from_pretrained( |
| | "KE-Team/KE-SemanticVAD").to('cuda') |
| | tokenizer = AutoTokenizer.from_pretrained("KE-Team/KE-SemanticVAD") |
| | |
| | |
| | # System Prompt |
| | AGNET_SPKING_SYS='# Role\n你是人机实时交互的**用户行为分析**模块，你将收到包含部分历史信息的 Human 和 Agent 最新实时对话记录 (Dialog)\n\n# 任务\n当前【Agent正在发言】，在此过程中，你需要基于对话分析 Human 的意图属于 <打断> 还是 <附和>\n\n# 输出\n不要有多余的分析，仅严格输出以下二者之一: <打断> 或 <附和>\n\n# 判断标准\n## <打断> 的情况\nHuman 行为: 试图抢夺话题主导权\n特征包括:\n- 提供新概念/词汇/判断（如命名、定性、对比）\n- 提出问题或异议\n- 引入与当前话题无关的新话题\n\n## <附和> 的情况\nHuman 行为: 赞同 Agent, 期望 Agent 继续说\n特征包括:\n- 使用零内容反馈（嗯/啊/对）\n- 机械重复 Agent 中的原词/同义词\n- 表达简单的确认或同意（如“是的”、“没错”）\n' |
| | HUMAN_SPKING_SYS = '# Role\n你是人机实时交互的**用户行为分析**模块，你将收到包含部分历史信息的 Human 和 Agent 最新实时对话记录 (Dialog)\n\n# 任务\n当前【Human正在发言】，你需要基于对话判断 Human 是否已经完成发言\n\n# 输出\n严格输出以下二者之一: <完成> 或 <未完>\n\n# 判断标准\n## <完成> 的情况\nHuman 发言语义完整，说话很可能已经结束\n- 发言包含完整命题（如明确提问/请求/结论）\n- 出现结束性标记词（"好了"/"你觉得呢？"）\n\n## <未完> 的情况\nHuman 发言语义不完整，仍然可能继续说话\n- 语句末尾含连接词（"而且"/"不过"/"然后"）\n- 用户发言中夹杂思考词（"呃..."/"嗯..."）\n' |
| | SYS_MAP = dict( |
| | agent = AGNET_SPKING_SYS, |
| | human = HUMAN_SPKING_SYS |
| | ) |
| | |
| | # Dialog Format |
| | def dia_format(x): |
| | cur_spk = x['speaker'] |
| | system = SYS_MAP[cur_spk] |
| | if cur_spk == 'agent': |
| | u1 = x['history']['query'] |
| | a1 = x['history']['answer'] |
| | u2 = x['query'] |
| | dialog = f"# Dialog\nHuman[历史]:{u1}\nAgent:[实时]:{a1}\nHuman[实时]:{u2}\n" |
| | elif cur_spk == 'human': |
| | if len(x['history']) <= 1: |
| | u2 = x['query'] |
| | dialog = f"# Dialog\nHuman[实时]:{u2}\n" |
| | else: |
| | u1 = x['history']['query'] |
| | a1 = x['history']['answer'] |
| | u2 = x['query'] |
| | dialog = f"# Dialog\nHuman[历史]:{u1}\nAgent:[历史]:{a1}\nHuman[实时]:{u2}\n" |
| | else: |
| | raise ValueError('current speaker should in agent or human') |
| | return [{'role': 'system', 'content':system}, {'role': 'user', 'content': dialog}] |
| | |
| | |
| | # 数据样例 |
| | example = { |
| | "speaker": "agent", |
| | "query": "那具体是怎么实现的？比如说，如", |
| | "history": { |
| | "query": "怎么把人工智能技术用在虚拟现实开发上呢？", |
| | "answer": "将人工智能技术应用到虚拟现实开发中，可以通过智能算法来提升用户体验，比如使用机器学习来创建更真实的虚拟角色" |
| | } |
| | } |
| | messages = dia_format(example) |
| | text = tokenizer.apply_chat_template( |
| | messages, |
| | tokenize=False, |
| | add_generation_prompt=True |
| | ) |
| | model_inputs = tokenizer([text], return_tensors="pt").to(model.device) |
| | generated_ids = model.generate( |
| | **model_inputs, |
| | max_new_tokens=64 |
| | ) |
| | generated_ids = [ |
| | output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids) |
| | ] |
| | response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] |
| | print(f"检测结果: {response}") # -> <打断> |
| | ``` |
| |
|
| | # Citation |
| |
|
| | Please cite our Hugging-Face when using our code, data or model. |
| | ```Bibtext |
| | @misc{KE-SemanticVAD, |
| | author = {KE-TEAM}, |
| | title = {KE-SemanticVAD}, |
| | year = {2025}, |
| | publisher = {Hugging Face}, |
| | url = {https://huggingface.co/KE-Team/KE-SemanticVAD} |
| | } |
| | ``` |
| |
|
