Abstarct: Automatic language identification consists in recognizing a language baxsed on a sample of speech from an unknown speaker. Automatic language identification can help relation between people of various areas. It has multiple usages in development of tourism, free trade, amplification of national security by means of pre-processing and filtering of doubtful conversations, emergency service and simultaneous translations in international congresses and conversations. In this thesis, the system of automatic language identification is designed and simulated by assisting of classifying of various features. Therefore, we find suitable features for every language and train the classifying algorithm baxsed on features which selected by benchmarking attribute algorithm and multi-interval discretisation for various languages. Then, we determine decision rules for each language and use this classification rules to identify the language of testing samples. We use the samples of 10 seconds and 45 seconds in the OGI_TS databaxse to test the proposed method. There are acoustic samples in 11 languages, consist of: English, Farsi, French, German, Hindi, Japanese, Korean, Mandarin, Spanish, Tamil, and Vietnamese. The utterances ranged in duration from one second to 50 seconds, with an average duration of 13.4 seconds. Language identification systems, usually, use 9 initial languages. Thus, we implement our experiments on these 9 languages, and compare our results with other reported methods. The experiments are accomplished baxsed on various features such as wavelet transforms, MFCC, PLP and LPC. So far, different approaches are proffered for automatic language identification which are dependent on phonotactics information and their utilization is difficult. In this research, we present an approach independent of phonotactics that it is very easy and can identify languages with a good accuracy rate. In this method, we utilize the wavelet transform and cepstral coefficients that are available on different languages which don't have any requirements to linguistic information. Cepstral coefficients achieve the accuracy rate better than wavelet transform. Also the samples of 45 seconds wavelet transform and cepstral coefficient have better