This study investigates the possibility to use automatic, multi-level features for the prediction of L2 speech proficiency. The method was applied on a corpus containing audio recordings and transcripts for 38 Japanese learners of French who participated in a semi-spontaneous oral production task. Each learner's speech proficiency level was assessed by three experienced French teachers. Audio recordings were processed to extract features related to the pronunciation skills and phonetic fluency of the learners, while the transcripts were used to measure their lexical, syntactic, and discursive abilities in French. A Lasso regression using a leave-one-out cross-validation procedure was used to select relevant features and to accurately predict speech proficiency scores. The results show that five features related to the phonetic fluency (speech rate), lexical abilities (lexical density), discourse planning and elaboration skills (number of hesitation and false starts, mean utterance length) of the learners can be used to predict speech proficiency ratings (r = 0.71, mean absolute error on a 5-point scale: 0.53).