Purpose: This meta-analysis study aimed to determine the optimal level of talker variability in training to maximize second-language speech learning. Method: We conducted a systematic search for studies comparing different levels of talker variability in nonnative speech training, published through July 2024. Two independent reviewers screened studies for eligibility, extracted data, and assessed the risk of bias. A Bayesian network meta-analysis was implemented to estimate relative effect sizes of different talker variability training conditions and rank these conditions by their posterior probabilities using surface under the cumulative ranking curve (SUCRA) values. Results: A total of 32 studies involving 998 participants were analyzed to compare six training conditions based on the number of talkers. Using a no-training control condition as the reference and excluding the outlier, the random-effects model showed that training with six talkers was most effective (SUCRA = 94%, standardized mean difference [SMD] = 2.09, 95% CrI [1.30, 2.89]), exhibiting moderate between-study heterogeneity (posterior median SD = 0.60, 95% CrI [0.39, 0.90]). However, when considering both the format of talker presentation and training exposure, the conditions with four talkers presented in blocks across training sessions (SUCRA = 77%, SMD = 1.47, 95% CrI [0.92, 2.10]), two talkers intermixed during sessions (SUCRA = 75%, SMD = 1.65, 95% CrI [0.24, 3.03]), and six talkers intermixed (SUCRA = 72%, SMD = 1.38, 95% CrI [0.97, 1.79]), all showed similarly high effectiveness with only minor differences. Conclusions: This systematic review and Bayesian network meta-analysis demonstrate for the first time that optimizing talker variability in nonnative speech training requires a careful balance between the number of talkers and the presentation format. The findings suggest that a moderate level of talker variability is most effective for improving second-language speech training outcomes.