I want to build a classifier for my problem statement and for that I don't have data. So while doing data acquisition, what should be the minimum sample size? And would it be a good practice if I label each observation myself to build a valid data set? (I cannot automate the process of labeling observation to each class while doing data acquisition and manually doing that takes up lot of time)
Unfortunately you're not going to be able to do much without at least 200-300 records. You're going to be limited to simple (i.e. mostly linear) models until your dataset expands to at least 1,000. Anything less than 1,000 will require very thorough cross validation, and if you're not careful you'll be at risk of building a model that easily overfits.
@EricLecoutre makes a great point that you should use Amazon's Mechanical Turk. It usually costs just a penny or two per record and could save you a lot of time.