Play
summary
Spotify voice personalization serves ~299 million users over 16 million tracks and 1.5 million podcast titles and treats voice as an audio to audio, zero GUI problem that requires getting the first result right. Pipeline is ASR → NLU (intent and slots) → fulfillment where candidates come from recommendation generators and a personalized ranker. Solutions include contextual multi-armed bandits to pick play contexts, a candidate stack with a personalized ranker, and epsilon exploration to collect unbiased data. Key pains are ASR errors, accent mis-transcriptions, and evaluating precision at one.
AI-written summary · about
chapters (22)
- 0:00Intro
- 1:13Personalization a Spotify
- 2:47Outline
- 3:17The Voice Experience
- 3:57Surfaces
- 4:17Utterance Type
- 6:11Utterance Trends
- 6:37Paradigm
- 8:53System Overview
- 9:53End to End
- 10:56Interesting Problems
- 12:38Play Artist
- 13:08No one size fits all
- 13:47Contextual MAB
- 14:26Descriptive
- 15:39Flavors of Recommendations
- 15:51Personalized Ranking
- 16:22AB Testing
- 17:23Open Challenges
- 20:22Future Work
- 20:54Team
- 24:22ML conf Online
related talks
19m‘The Voice’: New Challenges in a Zero UI world
40mPowering Spotify's Audio Personalization Platform
19mAudio is everywhere
20mPersonalization in the Era of LLMs
34mBuilding a Large Music Recommender Leveraging AI, Deep Learning and Human Expertise
39mIt's All in the Data: The ML Behind Alexa's AI Systems
28mLearning Fast: Home Experimentation and Open Research at Spotify
29mHow Spotify Does ML At Scale