I would recommend OCaml.
In my personal point of view, the main basis of modern¹ functional programmings are higher-order functions, a static type system, and algebraic datatypes and pattern matching.
Between a Scheme, a ML and a Haskell, I would choose the ML because I think it's the most relevant to this definition. Scheme doesn't have static typing (there is Typed Racket, but it's not for the scheme beginners), and Haskell has too much other stuff (monads, lazy evaluation...) that, while interesting, may divert attention from the important basis.
SML and OCaml are equally interesting; I'm more used to OCaml, and it has a more "practical" feeling that is nice (but if you really want "practical" to the risk of losing your soul, you may as well pick F#).
Scheme and Haskell are also very interesting languages. Scheme emphasis on macros is interesting, but not directly related to functional programming (it's another of the things in the world you should definitely try, as well as logic programming, stack-based languages, and capability-oriented E).
Haskell is definitely a great language and, I think, a mandatory point for the aspiring gurus of functional programming. But as the core languages of OCaml and Haskell are very much similar (except for lazy evaluation that is distracting for the beginner), it's easy to learn Haskell once you know the basics of OCaml. Or rather, you can concentrate on the weird stuff, and you don't have to assimilate the basics at the same time.
Similarly, once you have seen OCaml, and possibly also Haskell, and still want to learn more, you should look at Coq or Agda. Yet few would recommend Coq or Agda for a first introduction to functional programming...
To make my point clear : I think that learning OCaml (or SML) then Haskell will make you as good a functional programmer as learning Haskell directly, but more easily (or less painfully).
Besides, OCaml and Haskell both have good things differentiating them, and it's interesting to know about the advanced features of both. Just learning Haskell is poorer in that aspect (though of course you could learn OCaml after Haskell; I think it's less logical and will get you more frustrated).
For learning OCaml I would recommend Jason Hickey's book draft (PDF).
¹this definition is controversial. Some Scheme folks will claim static typing has nothing to do with functional programming. Some Haskell people will claim that their definition of purity ("what Haskell does, but no more") is a sine qua non condition for being a functional language. I agree to disagree.
Best Answer
CL type system is more expressive than the Haskell one, e.g., you can have a type
(or (integer 1 10) (integer 20 30))
for a value1,2,...9,10,20,21,...,30
.However, Lisp compilers does not force their understanding of type safety down your throat, so you can ignore their "notes" - at your own risk.
This means that you can write Haskell in Lisp (so to speak) by declaring all value types and carefully making sure that all the necessary types are inferred, but then it is easier to use Haskell in the first place.
Basically, if you want strong static typing, use Haskell or OCaml, if you want strong dynamic typing, use Lisp. If you want weak static typing, use C, if you want weak dynamic typing, use Perl/Python. Each path has its advantages (and zealots) and disadvantages (and detractors), so you will benefit from learning all of them.