AuraNote · 听纪
端侧离线 AI 会议纪要 — 实时转写 · 声纹分离 · 全程本地
macOS 14+
Swift 5.9
v1.2.1
隐私优先
App Store
产品简介
听纪（AuraNote） 是一款仅支持 macOS、完全本地离线运行的智能会议纪要应用。基于 FunASR（sherpa-onnx）语音识别引擎与 Pyannote 3.0 声纹分离技术，所有音频数据均在本地处理，无需联网，你的会议内容永远不会离开你的设备。
核心功能
🎙️ 实时流式转写
基于 Paraformer Online 的毫秒级流式语音转文字，录音过程中实时上屏，自动标点恢复。
📁 音频文件导入
支持 WAV / MP3 / M4A / FLAC 等多种格式，自动提取录制时间元数据，全自动离线分析流水线。
🧑🤝🧑 声纹分离
三级声纹分离：能量 VAD → Pyannote 3.0 Neural VAD → eres2net 512维声纹提取 + 增量聚类。多说话人自动标注，已知联系人自动匹配。
📚 声纹人脉库
本地 SQLite 持久化联系人声纹向量，越用越准。支持手动创建、编辑和管理联系人档案。
🤝 会议管理
会议元数据管理、同主题延续录音、完整历史记录持久化存储。
🔥 热词注入
内置热词预设管理器，录音前选择行业分类，ASR 识别时自动注入，大幅提升专业术语准确率。
⚙️ 转写审阅
审阅模式下支持手动编辑文本、调整说话人标签，展示会议元数据与参会人统计。
🏠 离线沙盒存储
全部数据存储于 App 沙盒 Documents 目录，绝不上传，杜绝隐私泄露。
技术架构
| 层级 | 技术 |
|---|
| UI 框架 | SwiftUI（macOS 14+） |
| 数据库 | SQLite（本地沙盒） |
| 实时 ASR | Paraformer Online（sherpa-onnx 流式） |
| 离线 ASR | Paraformer Offline（sherpa-onnx，token 级时间戳） |
| 声纹分割 | Pyannote 3.0 INT8（10ms 帧级） |
| 声纹提取 | eres2net（512 维） |
| 标点恢复 | CT-Transformer |
| VAD | Silero VAD + 振幅 VAD |
| 推理引擎 | ONNX Runtime（sherpa-onnx 封装） |
AI 模型清单
| 模型 | 大小 | 用途 |
|---|
| Paraformer | 227MB (int8) / 823MB (fp32) | 伪流式实时 + 离线高精度转写 |
| CT-Transformer | 72MB | 标点符号恢复 |
| Silero VAD | 1.7MB | 语音活动检测 |
| Pyannote 3.0 INT8 | 1.5MB | 帧级说话人切换检测 |
| eres2net | ~40MB | 512 维声纹向量提取 |
所有模型以 ONNX 格式内置于 App Bundle，通过 sherpa-onnx 调用，无需网络，无需外部依赖。
处理流水线
实时录音模式：
麦克风 → 增益/降噪 → 重采样(16kHz)
├─→ [流式 ASR] Paraformer Online → 实时文本
└─→ [VAD 切割] 攒批 8s → Pyannote + eres2net → 说话人标签
↓ 停止录音
[离线精分析] 全量音频 → Pyannote + eres2net + 增量聚类
↓ 声纹 chunk
Paraformer Offline → 高精度转写 → 声纹对齐 → 联系人匹配
音频导入模式：
音频文件 → 元数据提取 → AVAssetReader 解码 → 同下离线精分析流水线
隐私承诺
- 零网络传输：所有音频、文本、声纹数据仅在本地处理
- 沙盒存储：数据仅存储于 App 沙盒
- 麦克风权限：用户主动点击录音时才启用
- 详见 隐私政策
文档
AuraNote · 听纪
On-Device AI Meeting Minutes — Live Transcription · Voiceprint Diarization · 100% Local
macOS 14+
Swift 5.9
v1.2.1
Privacy First
App Store
Overview
AuraNote (听纪) is a macOS-only, fully offline, on-device intelligent meeting minutes app. Built on FunASR (sherpa-onnx) speech recognition and Pyannote 3.0 voiceprint diarization technology, all audio data is processed locally — no network required, your meeting content never leaves your device.
Core Features
🎙️ Real-Time Streaming ASR
Millisecond-level Paraformer Online streaming transcription with automatic punctuation restoration during recording.
📁 Audio File Import
Supports WAV / MP3 / M4A / FLAC and more. Auto-extracts recording metadata and triggers offline analysis pipeline.
🧑🤝🧑 Speaker Diarization
Three-stage pipeline: Energy VAD → Pyannote 3.0 Neural VAD → eres2net 512-dim embeddings + incremental clustering. Auto-labeling and known-contact matching.
📚 Voiceprint Library
Persistent local SQLite voiceprint storage. Improves with use. Full contact profile management.
🤝 Meeting Management
Metadata management, continuous recording for multi-session meetings, persistent history storage.
🔥 Hotword Injection
Built-in hotword preset manager with industry categories. Auto-injected during ASR for superior domain terminology accuracy.
⚙️ Transcript Review
Full editing capabilities in review mode. Adjust speaker labels, view metadata and participant statistics.
🏠 Offline Sandbox Storage
All data stored in App sandbox Documents directory. Never uploaded. Zero privacy leaks.
Tech Stack
| Layer | Technology |
|---|
| UI | SwiftUI (macOS 14+) |
| Database | SQLite (local sandbox) |
| Online ASR | Paraformer Online (sherpa-onnx streaming) |
| Offline ASR | Paraformer Offline (sherpa-onnx, token-level timestamps) |
| Segmentation | Pyannote 3.0 INT8 (10ms frame-level) |
| Embedding | eres2net (512-dim) |
| Punctuation | CT-Transformer |
| VAD | Silero VAD + Amplitude VAD |
| Inference | ONNX Runtime (via sherpa-onnx) |
AI Models
| Model | Size | Purpose |
|---|
| Paraformer | 227MB (int8) / 823MB (fp32) | Pseudo-streaming real-time + offline high-precision transcription |
| CT-Transformer | 72MB | Punctuation restoration |
| Silero VAD | 1.7MB | Voice activity detection |
| Pyannote 3.0 INT8 | 1.5MB | Frame-level speaker change detection |
| eres2net | ~40MB | 512-dim voiceprint embedding extraction |
All models are bundled as ONNX in the App Bundle via sherpa-onnx — no network, no external dependencies.
Processing Pipeline
Live Recording Mode:
Mic → Gain/Denoise → Resample (16kHz)
├─→ [Streaming ASR] Paraformer Online → Real-time text
└─→ [VAD Cuts] Batch 8s → Pyannote + eres2net → Speaker labels
↓ Stop Recording
[Offline Refinement] Full audio → Pyannote + eres2net + Clustering
↓ Voiceprint chunks
Paraformer Offline → High-precision → Alignment → Contact matching
File Import Mode:
Audio file → Metadata extraction → AVAssetReader decode → Same offline pipeline
Privacy Commitment
- Zero network transmission — all audio, text, and voiceprint data stays local
- Sandbox storage — data stored exclusively in App sandbox
- Microphone consent — only activated when user starts recording
- See Privacy Policy
Documentation