streaming-audio-memory

该 Skill 通过 BlueColumn API 接收边缘设备的音频分片，按设备命名空间持久化记忆；服务端使用 Whisper 转写并提取人物、地点、事件、动作和意图，支持幂等重试、按设备查询历史音频及 SSE 流式返回，并在静默或持续流后生成会话摘要。

bluecolumnconsulting-lgtm

热度 8.91

最近更新 2026-08-25 03:40:13

分类：知识依赖：第三方 API应用场景：数据处理 / 摘要总结 / 笔记整理 / 工作流

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add streaming-audio-memory

Metadata

name: streaming-audio-memory
description: Ingest live edge-device audio (car, doorbell, wearable, kiosk) into per-device streaming memory and recall what a device heard. Use when an agent streams audio chunks from devices, monitors ambient audio, or answers questions about a device's audio history. Requires a BlueColumn API key (bc_live_*).

Streaming Audio Memory — BlueColumn Skill

Stream audio chunks from any edge device into persistent, per-device memory. Each chunk is transcribed (Whisper large-v3), mined for entities and intent, and indexed under <namespace>_audio_streaming_<deviceId> for instant recall.

Setup

Read TOOLS.md or the platform secret store for the BlueColumn API key (bc_live_*). Base URL: https://xkjkwqbfvkswwdmbtndo.supabase.co/functions/v1

Ingest a chunk

curl -X POST .../streaming-audio \
  -H "Authorization: Bearer <key>" \
  -H "Content-Type: application/json" \
  -d '{"deviceId": "dashcam-01", "audio": "<base64>", "format": "wav",
       "sampleRate": 16000, "timestamp": 1756000000000,
       "idempotencyKey": "chunk_dashcam-01_1756000000000_a1b2c3d4e5f6"}'

Idempotency key format: chunk_<deviceId>_<epoch-ms>_<6-12 hex>. Same key + same body returns the cached response; same key + different body = 409. Retries never double-process.

Recall what a device heard

curl -X POST .../streaming-audio/query \
  -H "Authorization: Bearer <key>" \
  -H "Content-Type: application/json" \
  -d '{"deviceId": "dashcam-01", "query": "what happened near the warehouse?"}'

Add ?stream=1 (or Accept: text/event-stream) for SSE: sources → delta → done events.

Pipeline (per chunk)

Transcription — Groq Whisper large-v3
Entity extraction — people / locations / events / actions
Intent classification — query | command | notification | warning | conversation
Persist — indexed segment in the device namespace
Session summary — auto-summarized after 30s silence or 5 min of stream

Workflow

Buffer device audio client-side in 5–30s chunks (WAV/Opus/PCM/MP3)
POST each chunk with a fresh idempotency key
On user/environment questions, hit /streaming-audio/query for that device
Surface session_summary memories for long-horizon context