Explore OpenAI's Whisper, a groundbreaking speech-to-text model capable of transcribing and translating 97 languages. Learn about its weakly supervised encoder-decoder transformer architecture, trained on 680,000 hours of audio. Discover the model's implementation, fine-tuning process, and multitask capabilities. Delve into topics such as data quality, pipeline structure, generalization, overfitting prevention, and the impact of model size on performance. Gain insights into the weekly supervise technique and how mixing tasks contributes to the model's versatility.
Overview
Syllabus
Intro
What is Whisper
Example Implementation
Weekly supervise
Finetuning
Mixing Tasks
Data Quality
Model
Pipeline
Generalization
Overfitting
Model size
Multitask performance
Taught by
sentdex
