Fast Independent Vector Extraction by Iterative SINR Maximization
Abstract
We propose fast independent vector extraction (FIVE), a new algorithm that blindly extracts a single non-Gaussian source from a Gaussian background. The algorithm iteratively computes beamforming weights maximizing the signal-to-interference-and-noise ratio for an approximate noise covariance matrix. We demonstrate that this procedure minimizes the negative log-likelihood of the input data according to a well-defined probabilistic model. The minimization is carried out via the auxiliary function technique whereas, unlike related methods, the auxiliary function is globally minimized at every iteration. Numerical experiments are carried out to assess the performance of FIVE. We find that it is vastly superior to competing methods in terms of convergence speed, and has high potential for real-time applications.
Audio Samples
The algorithms are
- FIVE: The proposed method.
- OverIVA: Our previous method [Scheibler and Ono, WASPAA 2019]
- AuxIVA: Full blind source separation with AuxIVA [Ono, WASPAA 2011], selection of output source with largest power
The signal-to-interference and noise ratio is fixed to 5 dB. The source model is time-varying Gaussian.
| mics | sample | algo. | clean | mix | output | SDR | SIR | iter. | runtime |
| 2 | 1 | FIVE | 9.00 dB | 12.84 dB | 3 | 0.131 s | |||
| 2 | 1 | OverIVA | 8.95 dB | 12.84 dB | 10 | 0.234 s | |||
| 2 | 1 | AuxIVA | 9.20 dB | 13.21 dB | 50 | 2.047 s | |||
| 2 | 2 | FIVE | 4.93 dB | 8.99 dB | 3 | 0.198 s | |||
| 2 | 2 | OverIVA | 4.33 dB | 8.26 dB | 10 | 0.324 s | |||
| 2 | 2 | AuxIVA | 7.27 dB | 11.46 dB | 50 | 2.427 s | |||
| 2 | 3 | FIVE | 7.44 dB | 10.85 dB | 3 | 0.168 s | |||
| 2 | 3 | OverIVA | 7.22 dB | 10.44 dB | 10 | 0.230 s | |||
| 2 | 3 | AuxIVA | 7.30 dB | 10.64 dB | 50 | 1.625 s | |||
| 3 | 1 | FIVE | 12.09 dB | 18.01 dB | 3 | 0.217 s | |||
| 3 | 1 | OverIVA | 12.08 dB | 18.36 dB | 10 | 0.294 s | |||
| 3 | 1 | AuxIVA | 12.68 dB | 20.50 dB | 50 | 3.365 s | |||
| 3 | 2 | FIVE | 7.80 dB | 13.99 dB | 3 | 0.199 s | |||
| 3 | 2 | OverIVA | 6.83 dB | 12.86 dB | 10 | 0.366 s | |||
| 3 | 2 | AuxIVA | 9.78 dB | 18.10 dB | 50 | 3.145 s | |||
| 3 | 3 | FIVE | 11.05 dB | 17.21 dB | 3 | 0.236 s | |||
| 3 | 3 | OverIVA | 10.74 dB | 16.67 dB | 10 | 0.318 s | |||
| 3 | 3 | AuxIVA | 10.55 dB | 16.40 dB | 50 | 3.989 s | |||
| 5 | 1 | FIVE | 15.60 dB | 22.01 dB | 3 | 0.416 s | |||
| 5 | 1 | OverIVA | 15.72 dB | 23.63 dB | 10 | 0.574 s | |||
| 5 | 1 | AuxIVA | 16.29 dB | 30.70 dB | 50 | 9.467 s | |||
| 5 | 2 | FIVE | 10.94 dB | 19.25 dB | 3 | 0.374 s | |||
| 5 | 2 | OverIVA | 10.13 dB | 19.04 dB | 10 | 0.590 s | |||
| 5 | 2 | AuxIVA | 9.76 dB | 24.72 dB | 50 | 10.101 s | |||
| 5 | 3 | FIVE | 13.30 dB | 19.85 dB | 3 | 0.371 s | |||
| 5 | 3 | OverIVA | 12.86 dB | 19.71 dB | 10 | 0.641 s | |||
| 5 | 3 | AuxIVA | 13.97 dB | 22.14 dB | 50 | 9.901 s | |||
| 8 | 1 | FIVE | 16.49 dB | 23.89 dB | 3 | 0.633 s | |||
| 8 | 1 | OverIVA | 17.40 dB | 31.68 dB | 10 | 0.910 s | |||
| 8 | 1 | AuxIVA | 17.50 dB | 33.20 dB | 50 | 22.396 s | |||
| 8 | 2 | FIVE | 12.44 dB | 21.82 dB | 3 | 0.747 s | |||
| 8 | 2 | OverIVA | 12.51 dB | 23.40 dB | 10 | 1.129 s | |||
| 8 | 2 | AuxIVA | 10.48 dB | 23.61 dB | 50 | 22.919 s | |||
| 8 | 3 | FIVE | 14.03 dB | 21.15 dB | 3 | 0.655 s | |||
| 8 | 3 | OverIVA | 14.26 dB | 21.88 dB | 10 | 0.937 s | |||
| 8 | 3 | AuxIVA | 13.33 dB | 20.47 dB | 50 | 23.327 s |
