It's All in the Data: The ML Behind Alexa's AI Systems
summary
Alexa runs a cloud pipeline: wake‑word detection on device, ASR, NLU, skill selection and TTS, and hardware stays current because models update in the cloud. ASR uses active learning—semi‑supervised models send low‑confidence audio for human transcription. NLU moves from flat intents to a factored graph using AMRL (schema.org compatible). Skill selection uses a hypothesis ranker (bi‑LSTM shortlist + contextual rerank). India needs custom slot priors (movie→song), cricket intent boosts, and user prefs learn via song‑duration matrix factorization (listen >30s = +1, stop <30s = −1).
AI-written summary · about
chapters (14)
- 0:00Start
- 10:33Speech to Text
- 12:09ASR - Automatic Speech Recognition
- 13:44Speech Recognition - India
- 18:47Challenges
- 20:47Alexa Meaning Representation Language (AMRL)
- 22:44Alexa Schema Ontology
- 23:13Type Ambiguity
- 23:50Cross Domain Re Use
- 27:57Customizing NLU for India
- 30:03How is data used in Machine Learning?
- 32:21Intent Training: Not Enough Data
- 34:05Text to Speech
- 34:45Under the hood
