A deep tensor-based approach for automatic depression recognition from speech utterances