Objective Evaluation of Noise Reduction Algorithms in Speech Applications

Parsa V, Umapathy K. Objective Evaluation of Noise Reduction Algorithms in Speech Applications. In: AES Convention 115. Audio Engineering Society; 2003. Paper 5952. Available from: https://aes.org/publications/elibrary-page/?id=12383

@inproceedings{Parsa2003_12383,
  author = {Parsa, Vijay and Umapathy, Karthikeyan},
  title = {{Objective Evaluation of Noise Reduction Algorithms in Speech Applications}},
  booktitle = {AES Convention 115},
  note = {Paper 5952},
  year = {2003},
  month = oct,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=12383}
}

TY  - CPAPER
TI  - Objective Evaluation of Noise Reduction Algorithms in Speech Applications
AU  - Parsa, Vijay
AU  - Umapathy, Karthikeyan
T2  - AES Convention 115
M1  - Paper 5952
PY  - 2003
DA  - 2003/10/06
UR  - https://aes.org/publications/elibrary-page/?id=12383
PB  - Audio Engineering Society
LA  - en
AB  - We have evaluated objectively the comparative performance of five noise reduction algorithms. These algorithms were based on the Short-Time Spectral Amplitude (STSA) estimation, subspace projections, wavelet packets with auditory masking, and time-frequency decompositions using matching pursuits. Speech stimuli corrupted by speech-shaped noise and multi-talker babble at 5 different Signal-to-Noise Ratios (SNRs) were used to test the performance of the noise reduction algorithms. Noise reduction performance was quantified using two different methods. In the first method, the Perceptual Evaluation of Speech Quality (PESQ) measure was computed twice - once between the original and noisy speech and the other between the original and enhanced speech. The difference between these two PESQ values indicated the performance of the noise reduction algorithm. The second method was based on the "phase reversed noise" technique where the noise reduction algorithm was tested twice, once with speech + noise and then with speech + phase reversed noise. The PESQ and SNR gain measures were then computed on the combined output. The results obtained from this study indicate that the STSA based algorithm performs better in terms of the amount of noise reduction, while the wavelet packet based algorithm performs better in terms of minimizing the speech distortion introduced by the noise reduction process.
ER  -