In this paper, we compare four typical spoken language identification (LID) systems. We introduce a novel acoustic segment modeling approach for the LID system frontend. It is assumed that the overall sound characteristics of all spoken languages can be covered by a universal collection of acoustic segment models (ASMs) without imposing strict phonetic definitions. The ASM models are used to decode spoken utterances into strings of segment units in parallel phone recognition (PPR) and universal phone recognition (UPR) frontends. We also propose a novel approach to LID system backend design, where the statistics of ASMs and their co-occurrences are used to form ASM-derived feature vectors, in a
vector space modeling (VSM) approach, as opposed to the traditional language modeling (LM) approach, in order to discriminate between individual spoken languages. Four LID systems are built to evaluate the effects of two different frontends and two different backends. We evaluate the four systems based on the 1996, 2003 and 2005 NIST Language Recognition Evaluation (LRE) tasks. The results show that the proposed ASM-based VSM framework reduces the LID error rate quite significantly when compared with the widely-used parallel PRLM method. Among the four configurations, the PPR-VSM system demonstrates the best performance across all of the tasks.