The invention is suitable for the field of noise reduction and provides an adaptive noise reduction method based on audio feature extraction. The method comprises the following steps: A, collecting sound through a pickup; B, carrying out segmented analysis and processing of the collected sound according to a frequency range; C, comparing the processed sound with a preset frequency range and determining whether the sound is within the preset range, going on to execute the next step if the sound is within the preset range, otherwise, carrying out denoising processing; D, determining whether the sound within the preset frequency range is bigger than a preset value a of a physical sample, executing the next step if the sound is bigger than the preset value a, or determining that the sound is sound of human talk if the sound is smaller than the preset value a; E, determining whether the sound within the preset frequency range is bigger than an acoustic feature preset value b, executing the next step if the sound is bigger than the preset value b, or determining that the sound is sound of human talk if the sound is smaller than the preset value b; and F, determining whether the sound within the preset frequency range is bigger than a semantic feature preset value c, determining that the kept sound is sound of music if the sound is bigger than the preset value c, or determining that the sound is sound of human talk if the sound is smaller than the preset value c.