The present study investigated the existence of a Japanese mental syllabary and units stored therein for speech production. Experiment 1 compared naming latencies between high and low initial mora frequencies using CVCVCV nonwords, indicating that nonwords with a high initial mora frequency were named faster than those with a low frequency initial mora. Experiments 2 and 3 clarified the possibility of CV light and CVN/CVR heavy syllables as being units implicated in speech production. CVNCV nonwords in Experiment 2 and CVRCV nonwords in Experiment 3 displayed shorter naming latencies and lower error rates than their baseline (same bi-mora frequencies) of CVCVCV-structured nonwords. Since bi-mora frequencies between CVN/CVR and CVCV were the same, heavy syllables comprised of CVN and CVR units may contribute to ready-made motor-programs stored in the Japanese mental syllabary as variations of the 100 core light syllables (300 units in total). Experiment 4 further tested the effects of bi-mora frequency on the naming of nonwords, and found that CVCVCV-structured nonwords with high bi-mora frequencies were named more quickly and accurately than those with low bi-mora frequencies, although some bi-mora combinations seem to exhibit nonconforming tendencies (i.e., null significance in item analysis). Experiment 5 demonstrated that the naming of real words with high word frequency was quicker than for other real word conditions with low word frequencies (i.e., word frequency effects), with little effect of bi-mora frequencies. Unlike the nonword condition of Experiment 4, bi-mora frequency had only a minor influence on the naming of real words. Based on these findings, the present study proposes a possible model of the Japanese mental syllabary accompanied by a discussion of bi-mora and word frequency effects. (Contains 12 tables and 3 figures.)