Most voicebots still ignore, nowadays, user fluency level, although recognizing it would allow to give answers to adaptation issues, according to the language level of the interlocutor. Towards to such an end, this paper proposes a fluency classification model using a small audio dataset. We extract various features such as Mel-frequency cepstral coefficients (MFCC) from the audio. Using recent classification models, such as CNN-LSTM-Attention and Wav2vec 2.0, we propose fluency classification models through three usual categories (Low, Intermediate and High). Furthermore, we demonstrate that simple data augmentation methods can improve classification accuracy. We employ several simple data augmentation techniques, such as speed and pitch scale variation. This augmentation multiplies by 6 the number of training samples when applied only to original samples, and by 32 when also applied to augmented samples.