Sentiment analysis is an engrossing and intriguing area of research because of its extensive significance in different domains. Assembling opinions of people about products, social and political events and problems through the Web is becoming progressively popular day by day. The views of users are helpful for the public and stakeholders when making certain decisiveness. Automatic sentiment analysis for natural audio streams containing spontaneous speech is an ambitious area of research that has accepted little attention. Accordingly, efficient abstract methods are required for mining and summarizing the audio-text from corpuses which, requires knowledge of sentiment-bearing words in speech. Many computational techniques, models and algorithms are there for mining opinion components from unstructured text. In this study, we have used lexicon-based approach MNB classifier and deep learning approach for automatic recognition of sentiment from natural speech and compare their results.