Analysis of heart sound (HS) signal is a significant approach for detecting cardiovascular diseases (CVDs). Specifically, heart murmurs are regarded as the first indication of pathological occurrences and carry important diagnostic information. With the aids of computer and artificial intelligence technologies, a lot of HS analysis methods are suggested, which principally fall into two kinds: acoustic analysis and time-frequency analysis. However, most of existing methods are associated poorly with diagnostic information in heart murmurs, which restricts severely further interpretations. Aiming to handle this bottleneck problem, a novel enveloped-form heart murmur feature extraction methods is proposed, which extracts features merely and directly from heart murmurs. Initially, the influences of fundamental HSs are eliminated and the envelopes of heart murmurs are acquired, by employing discrete wavelet transform, Shannon envelope, as well as detecting and selecting peaks of heart murmurs. Thereafter, two key features SP and TS (the ratios of start position and time span of the envelopes of heart murmurs to the length of a HS cycle respectively) are extracted directly from the envelopes of heart murmurs, which are according to that the envelopes of different heart murmurs are of diverse shapes. By applying the key features to artificial neural network for classification and CVD diagnosis, the diagnostic accuracy is up to 96 %, which significantly validates the practicability and effectiveness of the proposed method.