This article introduces a new sentiment analysis framework that extracts information from audio records to predict stock returns. Unlike the previous scoring approach in finance literature that relies on the vocal emotion analysis software, this approach combines Speech Emotion Recognition and Transfer Learning to deliver customized sentiment scores for individual research purposes. It contains three components: feature extraction, sentence-level transfer learning, and scoring aggregation via penalized likelihood. In the empirical analysis, I study the earning calls corpus and examine the incremental informativeness and the corresponding market reaction. I show that this approach excels at extracting predictive signals and that speech sentiment is a remarkable predictor of abnormal returns. I then extract the text sentiment scores from the corresponding transcript using the BERT model and demonstrate that a trading strategy based on speech generates a Sharpe Ratio that is 2.75 times higher than the text strategy. Moreover, speech contains more information than text and a strategy combining both of them provides a Sharpe Ratio that is twice as large as the pure speech strategy. Eventually, information in speech is incorporated into prices with a delay thus it can be exploited in a real-time trading strategy.