The first protocol for password authenticated key exchange that appeared in the crypto community was the Bellovin-Merritt scheme (see also this survey page 4). This protocol is very simple, and might actually suit your need: is is exactly a Diffie-Hellman key exchange, in which the flows are encrypted with a block cipher (using the common password as the key of the cipher), and where the secret key the players agree on is derived by hashing the Diffie-Hellman tuple. The security of this protocol was analyzed several times, in various models (ideal cipher model or random oracle model, in indistinguishability-based framework or simulation-based framework...). Although it does not enjoy a proof of security in the plain model, you might be satisfied with a protocol proven secure in the random oracle model.
In this case, this scheme seems to exactly fit your requirements: you need "any Diffie-Hellman key exchanged", together with "any (good) hash function" and "any block cipher" that allows you to encrypt the flow with the password. Variations of the Bellovin-Merritt scheme that might make it even simpler are presented in this article (they essentially replace the block cipher by a simple one-time pad, and have two variants, one non-concurrently secure and one concurrently secure).
EDIT: so, after discussing with Ricky Demer, this does not quite work yet. A necessary condition for this to work is that the messages generated by the DHKE - which are group elements - should be indistinguishable from random bit-strings. For DHKE over $\mathbb{Z}^*_p$ for some large prime $p,$ group elements can be naturally mapped to a distribution statistically indistinguishable from bit-strings, and existing implementations might already encode these messages as random-looking bit strings (but this would have to be checked). For more sparse groups such as elliptic curves, I believe such a mapping can be done, but it would be more cumbersome from an implementation point of view. I thank Ricky Demer for pointing this out.
The variants presented in this article do not use a block cipher, which gives rise to dictionary attacks when the encoded element does not look random, but using a kind of multiplicative one-time pad: Alice masks her flow $g^x$ by multiplying it with $M^{\mathsf{pw}}$, where $M$ is a group element and $\mathsf{pw}$ is the password (and Bob plays similarly). Here, you do not have to care about how group elements are represented; however, you must perform an exponentiation (with a small exponent) and a multiplication, hence it does not makes a black-box use of the DHKE key exchange.
EDIT:
So, I discussed today with my PhD advisor, who happens to be the author of quite a number of papers in the PAKE area (in particular this paper which I had mentioned). It confirmed what I had started to think: it does not seem feasible to build a PAKE with a black box access to a DH key exchange and symmetric primitives. Somehow, you have to be at least able to multiply two group elements (hence you must know their structure). I cannot prove that it is infeasible, of course, but that is currently unknown in the scientific community, and not believed to be feasible.