Drop the Bass with Embedding and Vectors in Azure AI Search
summary
Demo shows building real vector search pipelines: ResNet18 image embeddings feed Azure AI Search indexes with 12,300 dog vectors and a 280,000-image auction training set that cuts prediction costs from €4,500 to €400 and is implemented in five days. Audio pipeline uses chroma-based “harmonic embeddings,” Recordbox metadata, ~632 sampled phrases and a 298,000-track Every Noise corpus, with beat/key normalization to 126 BPM and on-prem Spotify ANNOY for fast nearest-neighbor lookups enabling workable similar-sound retrieval and prototype live-mix generation.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 1:10Auction house to vector search
- 3:55Embeddings and cosine similarity
- 6:07Text embedding demo
- 11:09Keyword vs vector search
- 14:28When vectors outperform keywords
- 17:19Embedding math
- 18:47Image embedding and nearest neighbors
- 23:03ResNet for image embeddings
- 26:15Face recognition limits
- 30:53Music similarity and DJ mixing
- 37:52AI won’t replace musicians
- 40:02Finding similar music
- 43:58Camelot wheel mixing
- 47:29Extracting phrases with Recordbox
- 51:42DJ mixes with AI Search
- 56:01Music data challenges
- 58:07Every Noise dataset
- 1:00:40Live coding demo
- 1:02:47Q&A
related talks
01h 03mDrop the Bass with Embedding and Vectors in Azure AI Search
46mAdvanced RAG with Azure AI Search - top questions answered
15mAzure Cosmos DB as Vector Database
12mThe mechanics of integrating AI and vector search into Azure Database for PostgreSQL
44mSupercharge your data with Azure AI Search and OpenAI
50mBuilding AI ready applications
18m: Workforce Efficiency with AI-Driven Vector Search
44mVector Search Isn’t Enough