This paper describes a speech database developed for speech analysis and recognition research at ETL. Its system design concept, system configuration and phonetic labeling are presented as well as database statistics. The features of the system are 1) utterance text generation from a statistic-phonological view, and 2) fine acoustic-phonetic labeling by semiautomatic techniques. An acoustically compact segment called APSeg is adopted for the labeling of speech samples and acoustic-phonetic variations are represented by directed networks. Some sorted results of the labeling are shown for a real speech database of a phonemically balanced word set.