Whisper API 教程 2026：转录、翻译与会议智能

会议录音价值巨大，但大多数从未被回顾。本教程将构建一个自动化系统，对任何录制的会议进行转录、分析并提取可操作的情报。

Whisper 在 2026 年的能力

99+ 种语言，高准确率

专业词汇（医学、法律、技术）

多种音频格式：MP3、MP4、MPEG、MPGA、M4A、WAV、WEBM

词级时间戳

通过 API 提供 large-v3 模型

设置

python
from openai import OpenAI
from pathlib import Path
import json
client = OpenAI()

基础转录

python
def transcribe_audio(file_path: str, language: str = None) -> dict:
    with open(file_path, "rb") as audio_file:
        transcript = client.audio.transcriptions.create(
            model="whisper-1",
            file=audio_file,
            response_format="verbose_json",  # 包含时间戳和分段
            timestamp_granularities=["word", "segment"],
            language=language  # None = 自动检测
        )
    
    return {
        "text": transcript.text,
        "language": transcript.language,
        "duration": transcript.duration,
        "segments": transcript.segments,
        "words": transcript.words
    }
基本用法
result = transcribe_audio("meeting_recording.mp3")
print(f"语言: {result['language']}")
print(f"时长: {result['duration']:.0f}秒")
print(f"\n转录文本:\n{result['text'][:500]}...")

处理大文件：分块

python
from pydub import AudioSegment
import tempfile
import os
def transcribe_large_file(file_path: str, chunk_minutes: int = 10) -> str:
    """处理超过 25MB 的文件，通过分块实现。"""
    audio = AudioSegment.from_file(file_path)
    chunk_ms = chunk_minutes * 60 * 1000
    
    full_transcript = []
    
    for i in range(0, len(audio), chunk_ms):
        chunk = audio[i:i + chunk_ms]
        
        # 将块导出到临时文件
        with tempfile.NamedTemporaryFile(suffix=".mp3", delete=False) as tmp:
            chunk.export(tmp.name, format="mp3")
            
            try:
                result = transcribe_audio(tmp.name)
                full_transcript.append(result["text"])
            finally:
                os.unlink(tmp.name)
    
    return " ".join(full_transcript)

翻译（非英语转英语）

python
def translate_audio(file_path: str) -> str:
    """将任何语言转录并翻译成英语。"""
    with open(file_path, "rb") as audio_file:
        translation = client.audio.translations.create(
            model="whisper-1",
            file=audio_file
        )
    return translation.text
将西班牙语/法语/德语/日语等翻译成英语
english_text = translate_audio("spanish_meeting.mp3")
print(english_text)

会议智能系统

核心功能：将原始转录转化为可操作的会议情报。

python
import re
from dataclasses import dataclass
from typing import List, Optional
@dataclass
class MeetingInsights:
    summary: str
    action_items: List[dict]
    decisions_made: List[str]
    open_questions: List[str]
    attendees_mentioned: List[str]
    key_topics: List[str]
    sentiment: str
    follow_up_required: bool
MEETING_ANALYSIS_PROMPT = """分析此会议转录文本并提取结构化信息。
返回 JSON，格式为：
{
  "summary": "3-5 句执行摘要",
  "action_items": [
    {
      "task": "具体行动",
      "owner": "人员姓名或 'unassigned'",
      "due_date": "提到的日期或 null",
      "priority": "high/medium/low"
    }
  ],
  "decisions_made": ["决策 1", "决策 2"],
  "open_questions": ["问题 1", "问题 2"],
  "attendees_mentioned": ["姓名 1", "姓名 2"],
  "key_topics": ["主题 1", "主题 2"],
  "sentiment": "positive/neutral/negative/mixed",
  "follow_up_required": true/false
}
任务描述要具体。即使没有指定负责人，也要捕获所有行动项。"""
def analyze_meeting(transcript: str) -> MeetingInsights:
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {"role": "system", "content": MEETING_ANALYSIS_PROMPT},
            {"role": "user", "content": f"会议转录文本：\n\n{transcript}"}
        ],
        response_format={"type": "json_object"}
    )
    
    data = json.loads(response.choices[0].message.content)
    
    return MeetingInsights(
        summary=data.get("summary", ""),
        action_items=data.get("action_items", []),
        decisions_made=data.get("decisions_made", []),
        open_questions=data.get("open_questions", []),
        attendees_mentioned=data.get("attendees_mentioned", []),
        key_topics=data.get("key_topics", []),
        sentiment=data.get("sentiment", "neutral"),
        follow_up_required=data.get("follow_up_required", False)
    )

说话人分离

python
def identify_speakers(transcript: str, known_attendees: List[str] = None) -> str:
    """识别并标记转录文本中的不同说话人。"""
    attendee_context = ""
    if known_attendees:
        attendee_context = f"已知参会者：{', '.join(known_attendees)}"
    
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{
            "role": "user",
            "content": f"""识别此会议转录文本中的不同说话人。
            {attendee_context}
            
            格式为：
            [SPEAKER A]: 文本
            [SPEAKER B]: 文本
            
            如果能够根据上下文（介绍、提到的名字）识别说话人，请使用实际姓名。否则使用 Speaker A、B、C 等。
            
            转录文本：
            {transcript[:8000]}"""  # 截断以适应上下文窗口
        }]
    )
    
    return response.choices[0].message.content

完整流水线

python
def process_meeting(audio_path: str, attendees: List[str] = None) -> dict:
    print(f"正在处理：{audio_path}")
    
    # 步骤 1：转录
    print("  转录音频...")
    result = transcribe_audio(audio_path)
    transcript = result["text"]
    duration_minutes = result["duration"] / 60
    
    # 步骤 2：识别说话人
    print("  识别说话人...")
    labeled_transcript = identify_speakers(transcript, attendees)
    
    # 步骤 3：提取洞察
    print("  提取会议洞察...")
    insights = analyze_meeting(labeled_transcript)
    
    # 步骤 4：格式化输出
    output = {
        "file": audio_path,
        "duration_minutes": round(duration_minutes, 1),
        "language": result["language"],
        "transcript": labeled_transcript,
        "insights": {
            "summary": insights.summary,
            "action_items": insights.action_items,
            "decisions": insights.decisions_made,
            "open_questions": insights.open_questions,
            "key_topics": insights.key_topics,
            "sentiment": insights.sentiment
        }
    }
    
    # 步骤 5：保存结果
    output_path = Path(audio_path).stem + "_intelligence.json"
    with open(output_path, "w") as f:
        json.dump(output, f, indent=2)
    
    print(f"  完成！输出已保存至：{output_path}")
    print(f"  发现 {len(insights.action_items)} 个行动项")
    
    return output
用法
meeting_data = process_meeting(
    "q4_planning_meeting.mp3",
    attendees=["Sarah (CEO)", "Marcus (CTO)", "Priya (VP Sales)"]
)
打印摘要
print("\n=== 会议情报报告 ===")
print(f"时长：{meeting_data['duration_minutes']} 分钟")
print(f"\n摘要：\n{meeting_data['insights']['summary']}")
print(f"\n行动项：")
for item in meeting_data['insights']['action_items']:
    owner = item.get('owner', '未分配')
    due = item.get('due_date', '无日期')
    print(f"  [{item['priority'].upper()}] {item['task']} → {owner} ({due})")

与日历集成

python
import datetime
from googleapiclient.discovery import build
def post_to_google_calendar(meeting_data: dict, calendar_id: str, service):
    """将会议笔记添加到 Google 日历事件中。"""
    action_items_text = "\n".join([
        f"• {item['task']} ({item.get('owner', 'TBD')})"
        for item in meeting_data['insights']['action_items']
    ])
    
    description = f"""会议摘要
{meeting_data['insights']['summary']}
行动项
{action_items_text}
做出的决策
{chr(10).join(['• ' + d for d in meeting_data['insights']['decisions']])}"""
    
    # 查找今天的会议事件并更新描述
    now = datetime.datetime.utcnow()
    events_result = service.events().list(
        calendarId=calendar_id,
        timeMin=now.strftime("%Y-%m-%dT00:00:00Z"),
        maxResults=10,
        singleEvents=True
    ).execute()
    
    # 用会议笔记更新匹配的事件
    # 实现取决于如何将录音与日历事件匹配

成本与性能

文件长度转录时间API 成本

30 分钟会议~45 秒$0.27 60 分钟会议~90 秒$0.54 2 小时会议~3 分钟$1.08

分析成本 (GPT-4o)： 每次会议约 $0.10-0.30

每次会议总计： $0.40-1.40 —— 相比之下，人工转录每次会议 $10-20

结论

上述会议智能流水线将录制的会议转化为结构化的、可搜索的知识。每次会议成本低于 $1.50。按每周 10 次会议计算，每月 $60 即可确保不再丢失任何会议洞察。大多数团队发现仅行动项提取一项就足以证明成本的合理性——再也不用担心会议后谁负责什么了。