Objective. This study investigates the efficacy of electroencephalography (EEG) centered around the user's ears (ear-EEG) for a speech-imagery-based brain-computer interface (BCI) system. Approach. A wearable ear-EEG acquisition tool was developed and its performance was directly compared to that of a conventional 32-channel scalp-EEG setup in a multi-class speech imagery classification task. Riemannian tangent space projections of EEG covariance matrices were used as input features to a multi-layer extreme learning machine classifier. Ten subjects participated in an experiment consisting of six sessions spanning three days. The experiment involves imagining four speech commands ('Left,' 'Right,' 'Forward,' and 'Go back') and staying in a rest condition. Main results. The classification accuracy of our system is significantly above the chance level (20%). The classification result averaged across all ten subjects is 38.2% and 43.1% with a maximum (max) of 43.8% and 55.0% for ear-EEG and scalp-EEG, respectively. According to an analysis of variance, seven out of ten subjects show no significant difference between the performance of ear-EEG and scalp-EEG. Significance. To our knowledge, this is the first study that investigates the performance of ear-EEG in a speech-imagery-based BCI. The results indicate that ear-EEG has great potential as an alternative to the scalp-EEG acquisition method for speech-imagery monitoring. We believe that the merits and feasibility of both speech imagery and ear-EEG acquisition in the proposed system will accelerate the development of the BCI system for daily-life use.