Drop the Bass with Embedding and Vectors in Azure AI Search
summary
Live demos show Azure AI Search doing text, image, face and music vector search by indexing embeddings rather than keywords: uses Azure AI Search, ResNet-18 + PyTorch for 512‑dim image vectors, annoy for local ANN, librosa for chroma analysis and Sonic Pi for playback. A music dataset of ~450,000 MP3 samples yields ~300,000 embeddings (three weeks on a home PC); initial waveform-image matching sounds bad, but switching to chromatic (note‑based) embeddings plus tempo‑normalization to 126 BPM produces usable mixes and live demos. Face embedding demo works but hits legal limits.
AI-written summary · about
chapters (14)
- 0:00Embeddings and vectors in Azure AI Search
- 1:54Image embeddings to auction project
- 3:59Vector search and ANN theory
- 8:00Visualizing embeddings with UMAP
- 13:19Vector search in Azure AI Search
- 16:28Semantic search demo
- 23:08Image embeddings with CNNs
- 28:27Face embeddings demo
- 32:17Music embeddings DJ project
- 41:55Music theory Camelot wheel
- 45:30Track similarity chromatic embedding
- 53:00Mixing genres with Sonic Pi
- 56:15Music embeddings challenges
- 1:00:31Music coding resources
related talks
55mSupercharged Search with Semantic Search and Vector Embeddings
44mSupercharge your data with Azure AI Search and OpenAI
12mBuilding Generative AI Apps with Azure Cosmos DB
47mRAG at scale: production-ready GenAI apps with Azure AI Search
46mPower AI apps and develop rich experiences with Azure SQL Database
31mVue, You and AI Search
48mThe power of AI and Copilot for Azure Databases
18m: Workforce Efficiency with AI-Driven Vector Search