Computers are not yet capable of understanding speech in the same way as a human being. They do not have the same context or life experience that a human can bring to bear.
Speak & Improve uses a statistical analysis of a large number of features extracted from a learner's speech, which is then compared against the same features extracted from a large corpus of “training data” (speech from EFL students) for which we already have information about the level of proficiency.
These features are indicative of the student's level of attainment: some are indicative of good speaking, others of lower level production. Speak & Improve combines these positive and negative indications together to generate the final score for a learner's speaking.
Speak & Improve gives fairly accurate results across a wide range of learners’ submissions and, as this research project progresses and more learners use Speak & Improve, the accuracy of these scores will continue to improve.
However, this also means that Speak & Improve is only accurate on the type of speech that it’s been trained on. In particular, because it’s been trained on the speaking output of EFL learners, it does not provide accurate results for the speech of fluent native speakers. The features extracted from a fluent native speaker’s speech are different, and do not appear in the training data, so Speak & Improve is unable to judge them.
Simply put, your English may be “too English” for Speak & Improve to be able to judge it.
Given speaking output by genuine EFL students (the kind of speech that Speak & Improve has been trained on) you will see much more accurate results. If that is not the case, please do let us know.