In the steganography methods for low-bit-rate compressed speech based on Quantization Index Modulation (QIM), Nearest-neighbor Projection Point QIM (NPP-QIM) steganography has high embedding efficiency and security. Focusing on the issue that the accuracy of the existing steganalysis methods against the NPP-QIM steganography is not high, a steganalysis approach based on Bayesian inference was proposed for improving it. Firstly, Codeword Spatiotemporal Transition Network (CSTN) was constructed by using the Vector Quantization (VQ) codewords VQ1, VQ2, VQ3. Secondly, the codeword transition index was introduced to simplify the CSTN to obtain Steganography-Sensitive CSTN (SS-CSTN). Thirdly, Codeword Bayesian Network (CBN) was further constructed based on SS-CSTN. Finally, the network parameters of CBN were learned by utilizing Dirichlet distribution as the prior distribution to implement QIM steganalysis. The experimental results indicate that the detection accuracy of the proposed CBN method against the NPP-QIM steganography is improved by 25 percentage points and 37 percentage points compared with Index Distribution Characteristic (IDC) method and Derivative Mel-Frequency Cepstral Coefficients (DMFCC) method when the embedding strength is 100% and the speech length is 10 s. In the aspect of time performance, the CBN method can detect a 10 s speech segment in real time with about 21 ms.