Speech coding, audio signal processing, echo cancellation and source separation
When at Ericsson research during 1990-1995 work was performed in the following fields
- Low bit rate speech coding algorithms for the half-rate speech codecs for the GSM system and for the North-American D-AMPS system.
- Background sound coding algorithms.
- Echo cancellation algorithms, for network use and for acoustic applications in cellular phones.
- Statistical source separation algorithms.
The work on speech coding involved development and implementation of candidate half-rate speech-codecs for the North American TDMA system and for the GSM system, see the publication 11 for details. In particular a time variable spectral analysis method was developed, for improvement of the AR-modeling part of the speech codec. The method embeds a stepwise time variability in the model, using the AR-parameters of the previous frame and the estimated AR-parameters of the present frame as end values. Thereby it is sufficient to transmit the estimated parameters of the present frame to the receiving end for re-generation of the speech signal. The model is illustrated in Figure 1. Table 1 displays the obtained improvements.
Figure 1: Time variable AR-model.
Table 1: Measured performance gains in terms of segmental SNR obtained with time variable AR-modeling.
Audio quality problems caused by the codec used in the North American D-AMPS system was another field of reasearch. These problems were traced to an excessive non-stationarity of the AR-model part of the speech model employed by the VSELP speech codec.
In the echo control field, an Averaged Kalman Filter Algorithm (AKFA) was derived (1, 2). This algorithm provides optimal transient convergence of the adaptive FIR filter, as displayed in Figure 2. The computational complexity of the AKFA is the same as for the NLMS algorithm.
Statistical source separation algorithms exploiting dual microphones were also studied. The paper 9 presents a generalization of previous algorithms to the dynamic FIR filter case, and provides an initial analysis of the convergence properties of the generalized algorithm.
A new change detection scheme, based on an averaged low order model of the NLMS algorithm, was presented in 7.
Figure 2: Transient convergence of the time variant AKFA (fastest), the time invariant AKFA and the NLMS algorithm (slowest).
1. T. Wigren, Kalman filter echo canceller, US Patent 5 995 620 and International Patent , 1999.
2. T. Wigren, Fast converging and low complexity adaptive filtering using an averaged Kalman filter, IEEE Trans. Signal Processing , vol. 46, no. 2, pp. 515-518, 1998.
3. T. Wigren and A. Bergström, Transmission error concealment, US Patent 5598506 and International Patent , Jan. 28, 1997.
4. T. Wigren and A. Bergström, Lost frame concealment, US Patent 5596678 and International Patent, Jan. 21, 1997.
5. T. Wigren, A. Bergström and F. Jansson, Rejected frame concealment, US Patent 5572622 and International Patent , Nov. 5, 1996.
6. T. Wigren, Discriminating between stationary and non-stationary signals, US Patent 5579432 and International Patent , Nov. 26, 1996.
7. S. Halvarsson, T. Wigren and B. Wahlberg, A test statistics for low complexity change detection in dynamic systems based on averaged filter model, Proc. ICASSP , Atlanta, Georgia, U.S.A., pp. 2845-2848, 1996.
8. T. Wigren, A. Bergström, S. Harrysson, F. Jansson and H. Nilsson, Improvements of background sound coding in linear predictive speech coders, Proc. ICASSP, Detroit, Michigan, U.S.A., pp. 25-28, 1995.
9. U. Lindgren, T. Wigren and H. Broman, On local convergence of a class of blind separation algorithms, IEEE Trans. Signal Processing, vol. 43, no. 12, pp. 3054-3058, 1995
10. T. Wigren, Time variable spectral analysis based on iterpolation for speech coding, US Patent 5351388 and International Patent , Sep. 24, 1994.
11. T. B. Minde, T. Wigren, J. Ahlberg and H. Hermansson, Techniques for low bit rate speech coding using long analysis frames, Proc. ICASSP, Minneapolis, Minnesota, U.S.A., pt. II, pp. 604-607, 1993.