A number of task-based pronunciation teaching interventions have been shown to raise learners' awareness of relevant properties of L2 speech input during interaction (e.g., Solon et al., 2017). However, it still remains unclear whether pronunciation gains are generalisable to diverse lexical contexts (unfamiliar tokens) and elicitation modes (words vs. sentences). This study investigates whether a focus on phonetic form improves learners' production of English /iː – ɪ/ and /ae – ʌ/ and leads to generalisation effects and retention. Sixty-three L1 Catalan/Spanish EFL learners carried out 20 dyadic, problem-solving tasks over 7 weeks. Task completion required the distinction of the target lexical items (e.g., bean – bin, cat – cut). Gains in production and generalisation effects were assessed through delayed word and sentence repetition tasks, and Mahalanobis distances were measured between confusable vowels and between learners' and native speakers' productions. Results showed that not only did learners increase the qualitative distance between the target confusable vowels, but their L2 vowel productions also approximated those of native speakers at post-test in words produced in isolation and in sentences. In addition, gains generalised to untaught tokens and improvement was retained after 11 weeks.