Separation of two speakers recorded in an anechoic chamber

• Stereo recordings of two speakers were performed in our anechoic chamber. Separation is entangled by recording noise.

Experimental setup: The microphones were placed 35 cm apart. The speakers were standing in a distance of 3 m at 60 degrees to the left and at 60 degrees to the right of the mid-perpendicular of the microphones, respectively. A stereo recording of each speaker was performed (one channel of each original recording is shown above). The two stereo recordings were summed to obtain the mixed signals (both channels shown above).
Separation parameters: sampling rate 12 kHz, 256 samples long Hanning window, DFT length 512 samples, shift 64 samples.

• To compare our algorithm with other blind source separation algorithms, we have applied it to recordings by Te-Won Lee from a normal office room. The same data has also been separated by Lucas Parra.

Separation parameters: sampling rate 16 kHz, 384 samples long Hanning window, DFT length 512 samples, shift 64 samples.

• Another interesting test is this recording by Lucas Parra: In a reverberant environment a TV set interferes with a speaker's voice.