# Senseaudio Meeting Assistant

> 用于构建和排查 SenseAudio 会议助手，覆盖实时会议转写、说话人区分、实时翻译、会议纪要生成、行动项提取与转录导出。Build and troubleshoot SenseAudio meeting assistants for live meeting transcription, speaker-aware records, realtime translation, meeting note generation, action item extraction, and transcript export.

- Skill: `zrtch/senseaudio-meeting-assistant` (Agent Skill, multi-file: 8 files)
- Install (CLI): `npx skillmds@latest add zrtch/senseaudio-meeting-assistant`
- Raw SKILL.md: https://api.skillmd.com/api/skills/zrtch/senseaudio-meeting-assistant/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: zrtch (https://skillmd.com/u/zrtch)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/zrtch/senseaudio-meeting-assistant

---


# SenseAudio 会议助手 / SenseAudio Meeting Assistant

使用这个 skill 来构建基于 SenseAudio ASR 的会议类工作流。
Use this skill to build meeting-focused workflows on top of SenseAudio ASR.

## 先读哪些文件 / Read First

- `references/realtime-workflow.md` - 实时会议的 WebSocket 转写流程 / realtime WebSocket transcription flow for live meetings
- `references/offline-meeting-transcription.md` - 录音文件的 HTTP 转写、说话人分离、时间戳与翻译 / HTTP transcription with diarization, timestamps, sentiment, and translation
- `references/note-generation.md` - 如何把转写结果整理成会议纪要、决策与行动项 / how to turn transcripts into structured meeting notes, decisions, and action items
- `references/examples.md` - 常见会议助手场景的最小实现示例 / minimal examples for common meeting assistant tasks

按用户需求只加载必要的参考文件，不要一次性全读。
Load only the reference file(s) needed for the user's request.

## 适用场景 / When To Use

当用户需要以下能力时，使用这个 skill：
Use this skill when the user wants any of the following:

- 实时会议字幕或实时转写 / live meeting captions or realtime transcripts
- 带说话人区分的会议记录 / speaker-separated meeting records
- 双语会议输出或会中翻译 / bilingual or translated meeting output
- 会议摘要、决策提取、行动项整理 / meeting summaries, decisions, and action items
- 基于会议音频或识别历史构建可搜索归档 / searchable archives built from meeting audio or ASR history
- 将 SenseAudio ASR 与 LLM 摘要能力结合的会议助手 / a meeting copilot combining SenseAudio ASR with LLM summarization

## 能力路由 / Capability Routing

优先选择满足需求的最小实现。
Choose the smallest implementation that satisfies the request.

1. **实时会议助手 / Live meeting assistant** → 读取 `references/realtime-workflow.md`
2. **录音会议处理 / Recorded meeting processing** → 读取 `references/offline-meeting-transcription.md`
3. **音质预检 / Audio quality triage** → 先走音频质量分析接口
4. **历史记录查询 / Historical transcript lookup** → 使用识别记录查询接口
5. **纪要与行动项整理 / Structured notes and actions** → 再读取 `references/note-generation.md`

## 默认工作流 / Default Workflow

1. 先判断会议模式 / Pick meeting mode:
- 直播中 / 开会中 → 使用 WebSocket 实时识别 / live meeting → WebSocket realtime ASR
- 已有录音文件 → 使用 HTTP 文件转写 / uploaded recording → HTTP transcription

2. 再选择识别模型 / Choose recognition settings:
- `sense-asr-pro`：默认首选，适合会议场景，支持说话人分离、时间戳、情感与翻译 / best default for meetings needing diarization, timestamps, sentiment, and translation
- `sense-asr`：适合需要翻译但对会议增强能力要求较轻的场景 / useful when translation is needed but meeting constraints are lighter
- `sense-asr-deepthink`：适合实时 WebSocket 或更偏翻译理解的场景 / suitable for realtime WebSocket or translation-heavy semantic cleanup

3. 先构造最小可用请求 / Build a minimal valid request first:
- 加上 `Authorization: Bearer <SENSEAUDIO_API_KEY>`
- 先只保留必填参数 / keep only required fields initially
- 只有在确实需要时再开启说话人分离、时间戳、翻译、情感分析 / add diarization, timestamps, translation, or sentiment only when needed

4. 统一转写结果结构 / Normalize transcript output:
- 保留说话人、时间戳、翻译文本、分段边界 / preserve speaker, timestamps, translation, and segment boundaries
- 先把 ASR 原始结果归一化，再进入纪要或行动项生成 / normalize ASR output before summarization

5. 最后生成会议产物 / Generate meeting artifacts:
- 会议摘要 / summary
- 关键决策 / decisions
- 带负责人和时间定位的行动项 / action items with owner and timestamp when possible
- 可导出的 Markdown 或 JSON / export-ready Markdown or JSON

## 实现规则 / Implementation Rules

- 严禁硬编码密钥，统一从环境变量 `SENSEAUDIO_API_KEY` 读取。Do not hardcode secrets; read `SENSEAUDIO_API_KEY` from environment variables.
- 实时会议流式识别必须满足 `pcm`、`16000Hz`、单声道、16-bit little-endian。For realtime streaming, enforce `pcm`, `16000Hz`, mono, signed 16-bit little-endian.
- 录音会议优先使用 `response_format=verbose_json`，保留会议所需元数据。For recorded meetings, prefer `response_format=verbose_json`.
- 说话人分离只在支持的模型上使用，`max_speakers` 仅适用于 `sense-asr-pro`。Use diarization only with supported models; `max_speakers` is only for `sense-asr-pro`.
- 使用 `sense-asr-deepthink` 时不要传 `language`，只使用 `target_language`。Do not send `language` with `sense-asr-deepthink`; use `target_language` only.
- 会议纪要生成属于第二阶段，必须建立在结构化转写结果之上。Treat note generation as a second stage after transcript normalization.
- 如果用户要生产级代码，输出中要包含超时、重试、日志与失败处理。Include timeout, retry, logging, and failure handling for production code.

## 输出检查清单 / Output Checklist

当你给出实现方案时，至少包含：
When producing implementation output, include:

1. 选用的接口与模型 / chosen endpoint and model
2. 最小可运行请求示例 / minimal request example
3. 转写结果解析方式 / transcript parsing strategy
4. 纪要 / 行动项生成方式（如用户需要） / note and action-item generation strategy if requested
5. 会议场景下的限制说明，例如音频格式、说话人分离限制、翻译行为 / meeting-specific caveats such as audio format, diarization limits, or translation behavior

## 常见坑点 / Common Pitfalls

- **没有说话人字段 / Missing diarization fields**：确认已开启 `enable_speaker_diarization=true`，并使用 `sense-asr` 或 `sense-asr-pro`
- **没有字级时间戳 / No word timestamps**：确认设置了 `timestamp_granularities[]=word`
- **实时识别报错 / Realtime failures**：检查发送的是原始 PCM 数据，而不是 `wav/mp3` 容器字节
- **DeepThink 参数报错 / DeepThink errors**：移除不支持的 `language` 或其他不兼容参数
- **纪要质量差 / Weak meeting notes**：优先基于结构化 `segments` 总结，不要只对整段纯文本做摘要