Files:

Speech_AnechoicSpeechSignal.wav:
Anechoic speech signal from the TIMIT database, down-sampled from 16kHz to 4kHz. No changes made otherwise.

Speech_ReverberantSignal.wav:
The speech signal was distorted by WGN varying according to a random walk and normalised to 35dB SNR. The resulting noisy signal was filtered by the 8th order all-pole model of the gramophone horn response. Notice the metallic sound as well as the noise towards the end of the speech sequence. Segmental SRR:  -6.6264dB

Speech_TVAREstimate.wav:
The speech estimate acquired by the RBPF using the Markov chain based model with a model order of Q = 15 and 1000 particles. Notice that the signal quality was improved, however minor metallic effects are still audible. Segmental SRR: +2.9578dB (improvement of 9.5842dB compared to the reverberant signal)

Speech_PFSEstimate.wav:
The speech estimate acquired by the RBPF using the PFS model with 6 resonators and 1000 particles. Notice that the signal quality is significantly improved as compared to the reverberant signal. Towards the end of the sequence, minor noise effects are audible. Segmental SNR: +5.8113dB (improvement of 12.4377dB compared to the reverberant signal and 2.8535dB compared to the TVAR estimate).

Speech_PFSBeierholmEstimate.wav:
The speech estimate acquired by the RBPF using the PFS model by Beierholm and Winther with 6 resonators and 1000 particles. Segmental SNR: +2.123dB.

Speech_YegnanarayanaEstimate.wav:
The speech estimate acquired using the LPC residual approach by Yegnanarayana and Satyanarayana Murthy. Note that the metallic effects of the channel are still audible. Segmental SNR: 2.97dB. 