|
|
--- |
|
|
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} |
|
|
} |
|
|
``` |
|
|
|
