This dissertation investigates the soft typology of phonological patterns. By studying the types of patterns observed in the world's languages, we can do more than determine what is attested and what is unattested, or what is possible and what is impossible, but discover a wide range of more gradient phenomena. Phonological patterns range from being very common, observed in many language families across the continents of the world, to incredibly rare, known to be attested in only one language, to anywhere in between. These skews in typology, or soft biases, are challenging to explain via generative grammatical theories alone, like Optimality Theory Prince & Smolensky (1993/2004) and its descendants. Instead, in this dissertation, I argue that the differences in rate of attestation of phonological patterns can be understood in terms of differences in learnability across phonological patterns. Patterns that are relatively easy to learn are learned accurately and quickly, whereas harder to learn patterns tend to be learned less accurately and require more training data. These differences in learnability expand when applied to the generational stability model used in this dissertation, where artificial learners are exposed to a finite amount of data, after which they train a new generation of learners on whatever pattern they managed to learn. Harder to learn patterns are more likely to change across generations in such a model, and as a result, become less frequently attested in the languages of the world. In particular, this dissertation focuses on asymmetries in place of articulation contrasts of obstruent stops in word-initial and word-final positions. Chapter 2 investigates the soft typology of this domain through large scale typological surveys, and shows two major soft typological skews: The all-or-nothing skew, where languages are shown to be more likely to have either all of or none of [p t k] available in word-final position, rather than some subset, and the positional priority skew, which finds that languages are more likely to restrict a position: i.e. ban final stops, than a place of articulation: i.e. ban velar stops. Chapter 3 shows that these two typological skews are best captured through the interaction of grammar and the learning algorithm. Both of these skews are emergent predictions of the generational stability model with a set of constraints encoding markedness hierarchies that are well motivated from other phonological domains. The underattested patterns are shown to be harder to learn and relatively unstable across generations, and the common patterns are shown to be easier to learn and more stable. I compare these predictions to models of soft typology that prioritize the contribution of either grammatical structure or the learning algorithm, and show that neither approach captures the typology as well as an approach that focuses on the interaction of learning and grammatical structure. Chapter 4 investigates how the frequency of certain phonological forms within the lexicon of a language can contribute to the learnability of different patterns. Particular lexical frequency distributions can subvert general learnability patterns, potentially making rare patterns easier-to-learn than more common patterns. I investigate and characterize the properties of lexical frequency distributions that affect the learning of MaxEnt grammars. I then show how these properties can explain how the lexical frequency distributions of languages like Finnish can allow the otherwise rare and hard-to-learn phonological patterns exhibited in such languages to be learned stably across many generations. [The dissertation citations contained here are published with the permission of ProQuest LLC. Further reproduction is prohibited without permission. Copies of dissertations may be obtained by Telephone (800) 1-800-521-0600. Web page: http://bibliotheek.ehb.be:2222/en-US/products/dissertations/individuals.shtml.]