Linear Prediction Based Online Dereverberation and Noise Reduction Using Alternating Kalman Filters

S. Braun and E. A. P. Habets

IEEE/ACM Transactions on Audio, Speech and Language Processing, 2018.

Note: Please use the Chrome or Safari browser for optimal audio/video synchronization. If you encounter playback problems, please reload the page.

Example 1: Walking front-back

  • The signals were recorded using 2 omnidirectional microphones with 6 cm spacing.
  • The stationary noise PSD is estimated in advance during speech absence.
specgram_25_input

Example 2: Walking Walking left-right

  • The signals were recorded using 2 omnidirectional microphones with 6 cm spacing.
  • The stationary noise PSD is estimated in advance during speech absence.
specgram_28_input

Example 3: Babble noise

  • The signals in example 3 were generated using measured RIRs from a 2-channel microphone array with 20 cm spacing in a room with 0.7 s reverberation time.
  • The reverberated speakers (German and French) are located at different positions in the room.
  • Non-stationary babble noise recorded in a cafeteria was added to the reverberant signals with input SNR = 10 dB.
  • The noise covariance was estimated as a stationary average over periods of speech absence. Therefore, only the stationary part of the noise can be reduced, while non-stationary elements remain in the noise residual.
  • You can choose between different settings for the noise reduction (NR) and reverberation reduction (RR), which can be controlled independently for specific requirements or to subjective taste.
  • For these examples, the 2-channel output is played back, so you can localize the speakers on a stereo playback device.
Room3_babble_specgram_input