Abstract: Methods and apparatuses are described for generating a recommended conversation topic flow. A server executes a topic modelling algorithm on voice call transcripts to determine a distribution of topics present in the transcripts. The server generates a historical topic flow for each transcript by determining topic-related words in text blocks in the transcript and assigning a score to each text block. The server categorizes each topic flow as positive or negative based upon outcomes associated with the transcript. The server trains a classification model using the categorized topic flows as input. The server captures a bitstream corresponding to a current voice call and converts the bitstream into unstructured text. The server executes the classification model on the unstructured text to determine whether a call topic flow is positive or negative. The server generates topic recommendations and transmits the recommendations to a second client device.