100行代码构建智能会议纪要生成器：MobileBERT实战指南

100行代码构建智能会议纪要生成器：MobileBERT实战指南

【免费下载链接】mobilebert_uncased MobileBERT is a thin version of BERT_LARGE, while equipped with bottleneck structures and a carefully designed balance between self-attentions and feed-forward networks. 项目地址: https://ai.gitcode.com/openMind/mobilebert_uncased

引言：告别低效会议记录

你是否还在为冗长会议后的纪要整理焦头烂额？是否曾因遗漏关键决策而导致项目延误？本教程将带你用MobileBERT模型构建一个智能会议纪要生成器，仅需100行代码即可实现会议语音转写、关键信息提取和结构化纪要生成的全流程。

读完本文后，你将能够：

• 掌握MobileBERT模型的本地化部署方法
• 实现语音到文本的实时转换
• 提取会议中的决策、任务和时间节点
• 生成符合企业标准的结构化会议纪要

技术选型：为什么选择MobileBERT？

MobileBERT是Google于2020年推出的轻量级BERT模型，通过瓶颈结构和精心设计的注意力机制与前馈网络平衡，在保持BERT_LARGE 95%性能的同时，模型大小减少4倍，推理速度提升5倍。

环境准备：5分钟快速搭建

硬件要求

• CPU: 双核以上处理器
• 内存: 至少4GB RAM
• 存储: 1GB空闲空间（模型文件大小约400MB）

软件环境

# 克隆项目仓库
git clone https://gitcode.com/openMind/mobilebert_uncased

# 安装依赖
cd mobilebert_uncased
pip install -r examples/requirements.txt
pip install torch==2.8.0 transformers soundfile

核心功能实现：四大模块解析

1. 语音转文本模块

import soundfile as sf
import torch
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC

def speech_to_text(audio_path):
    # 加载预训练模型和处理器
    processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
    model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")
    
    # 读取音频文件
    speech, samplerate = sf.read(audio_path)
    
    # 处理音频并生成转录文本
    input_values = processor(speech, sampling_rate=samplerate, return_tensors="pt").input_values
    logits = model(input_values).logits
    predicted_ids = torch.argmax(logits, dim=-1)
    transcription = processor.decode(predicted_ids[0])
    
    return transcription

2. MobileBERT模型加载

from transformers import MobileBertTokenizer, MobileBertForQuestionAnswering
import torch

class MobileBERTProcessor:
    def __init__(self, model_path="./"):
        # 加载模型和分词器
        self.tokenizer = MobileBertTokenizer.from_pretrained(model_path)
        self.model = MobileBertForQuestionAnswering.from_pretrained(model_path)
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.model.to(self.device)
        
    def extract_info(self, context, question):
        # 编码输入
        inputs = self.tokenizer(question, context, return_tensors="pt").to(self.device)
        
        # 模型推理
        with torch.no_grad():
            outputs = self.model(**inputs)
            
        # 提取答案
        answer_start_index = torch.argmax(outputs.start_logits)
        answer_end_index = torch.argmax(outputs.end_logits) + 1
        
        # 解码答案
        predict_answer_tokens = inputs.input_ids[0, answer_start_index:answer_end_index]
        return self.tokenizer.decode(predict_answer_tokens)

3. 关键信息提取模块

def extract_meeting_info(transcript):
    processor = MobileBERTProcessor()
    
    # 定义需要提取的信息类型
    questions = {
        "decision": "What decisions were made?",
        "action_items": "What action items were assigned?",
        "deadlines": "What deadlines were mentioned?",
        "attendees": "Who attended the meeting?"
    }
    
    # 提取信息
    results = {}
    for key, question in questions.items():
        results[key] = processor.extract_info(transcript, question)
        
    return results

4. 纪要生成模块

def generate_minutes(meeting_info):
    """生成结构化会议纪要"""
    minutes = "# 会议纪要\n\n"
    
    # 添加基本信息
    minutes += "## 基本信息\n"
    minutes += f"- **日期**: {datetime.now().strftime('%Y-%m-%d')}\n"
    minutes += f"- **参会人员**: {meeting_info['attendees']}\n\n"
    
    # 添加决策事项
    minutes += "## 决策事项\n"
    minutes += f"- {meeting_info['decision']}\n\n"
    
    # 添加行动项
    minutes += "## 行动项\n"
    minutes += f"- {meeting_info['action_items']}\n\n"
    
    # 添加截止日期
    minutes += "## 截止日期\n"
    minutes += f"- {meeting_info['deadlines']}\n"
    
    return minutes

完整工作流：从语音到纪要

完整代码实现：

import soundfile as sf
import torch
import datetime
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC, MobileBertTokenizer, MobileBertForQuestionAnswering

class MeetingMinuteGenerator:
    def __init__(self):
        # 初始化语音转文本模型
        self.speech_processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
        self.speech_model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")
        
        # 初始化MobileBERT模型
        self.bert_tokenizer = MobileBertTokenizer.from_pretrained("./")
        self.bert_model = MobileBertForQuestionAnswering.from_pretrained("./")
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.bert_model.to(self.device)
        
    def speech_to_text(self, audio_path):
        # 语音转文本实现
        speech, samplerate = sf.read(audio_path)
        input_values = self.speech_processor(speech, sampling_rate=samplerate, return_tensors="pt").input_values
        logits = self.speech_model(input_values).logits
        predicted_ids = torch.argmax(logits, dim=-1)
        return self.speech_processor.decode(predicted_ids[0])
        
    def extract_info(self, context, question):
        # 信息提取实现
        inputs = self.bert_tokenizer(question, context, return_tensors="pt").to(self.device)
        with torch.no_grad():
            outputs = self.bert_model(**inputs)
        answer_start_index = torch.argmax(outputs.start_logits)
        answer_end_index = torch.argmax(outputs.end_logits) + 1
        predict_answer_tokens = inputs.input_ids[0, answer_start_index:answer_end_index]
        return self.bert_tokenizer.decode(predict_answer_tokens)
        
    def process_meeting(self, audio_path):
        # 完整处理流程
        transcript = self.speech_to_text(audio_path)
        
        questions = {
            "decision": "What decisions were made?",
            "action_items": "What action items were assigned?",
            "deadlines": "What deadlines were mentioned?",
            "attendees": "Who attended the meeting?"
        }
        
        results = {}
        for key, question in questions.items():
            results[key] = self.extract_info(transcript, question)
            
        return self.generate_minutes(results)
        
    def generate_minutes(self, meeting_info):
        # 纪要生成实现
        minutes = "# 会议纪要\n\n"
        minutes += "## 基本信息\n"
        minutes += f"- **日期**: {datetime.now().strftime('%Y-%m-%d')}\n"
        minutes += f"- **参会人员**: {meeting_info['attendees']}\n\n"
        minutes += "## 决策事项\n"
        minutes += f"- {meeting_info['decision']}\n\n"
        minutes += "## 行动项\n"
        minutes += f"- {meeting_info['action_items']}\n\n"
        minutes += "## 截止日期\n"
        minutes += f"- {meeting_info['deadlines']}\n"
        return minutes

# 使用示例
if __name__ == "__main__":
    generator = MeetingMinuteGenerator()
    minutes = generator.process_meeting("meeting_recording.wav")
    with open("meeting_minutes.md", "w") as f:
        f.write(minutes)
    print("会议纪要生成完成：meeting_minutes.md")

优化建议：提升性能与体验

模型优化

1. 量化处理：使用INT8量化减少模型大小和推理时间

from torch.quantization import quantize_dynamic
model_quantized = quantize_dynamic(bert_model, {torch.nn.Linear}, dtype=torch.qint8)

2. 模型缓存：首次运行后缓存模型文件

import os
from transformers import MobileBertTokenizer

CACHE_DIR = "./model_cache"
if not os.path.exists(CACHE_DIR):
    os.makedirs(CACHE_DIR)
    
tokenizer = MobileBertTokenizer.from_pretrained("./", cache_dir=CACHE_DIR)

功能扩展

1. 多语言支持：添加中文处理能力
2. 情感分析：识别会议中的情绪变化
3. 任务分配：自动关联任务到项目管理系统

结语：重新定义会议效率

本教程展示了如何利用MobileBERT模型构建一个高效的会议纪要生成器。通过将语音识别与自然语言处理相结合，我们实现了从会议录音到结构化纪要的全自动化流程。这个仅100行代码的解决方案不仅可以节省大量时间，还能确保会议信息的准确性和完整性。

随着远程办公的普及，这样的工具将成为团队协作的重要助手。下一步，你可以考虑添加实时会议转录功能或与日历系统集成，进一步提升工作效率。

最后，欢迎大家贡献代码和想法，一起完善这个项目！

【免费下载链接】mobilebert_uncased MobileBERT is a thin version of BERT_LARGE, while equipped with bottleneck structures and a carefully designed balance between self-attentions and feed-forward networks. 项目地址: https://ai.gitcode.com/openMind/mobilebert_uncased