In text-independent speaker identification, there are a large number of likelihood computations, especial in large population. To speed up the recognition, we proposed a lightweight algorithm called CBF (Codebook Filtering). CBF provides two phase of speaker pruning to accelerate the speaker recognition. To make CBF could process large population, this paper implements CBF on Map-Reduce framework. In this approach, we encountered some problems, such as how to balance accuracy and speed-up of algorithm. This paper provides a mechanism of parameter consulting to archieve satisfactory accuracy and speed-up factor. To verify algorithm, we implement it on Phoenix, a Map-Reduce framework on multi-core. As the result of experiment, this approach has increased the speed-up factor of CBF obviously. The speed-up factor reaches 40.2 when the accuracy keeps 94.98%.