Hello, I am doing some work in Speaker Recognition on GSM coded speech. Currently, I am doing experiments with the GSM EFR coder. My speech database is sampled at 16 kHz. So, to obtain signal sampled at 8 kHz to be fed into the coder, I have so far down-sampled using a 158th-order linear-phase FIR half-band filter, with a transition band of 150 Hz, a passband ripple of less than 0.1 dB, and more than 97 dB of attenuation in the stop band. Particularly I am interested in the speech path that goes from the Mobile Station (MS) to the PSTN (Public Switched Telephone Network). So I would like to have a filter more representative actual anti-aliasing low pass filter inside a mobile phone. Can anybody help me? The information I have found so far is: In GSM 06.51 (Enhanced Full Rate (EFR) speech processing functions; General description) Figure 1 /Transmit side, there is a block "MS side only", that shows the Low Pass Filter and the ADC, and refers to normative GSM 03.50. But, in the GSM 03.50 normative, the only place were the ADC is mentioned/shown is in figure 2, the analogue signal is coming either from hands-free, handset or headset MSs, and a note says that the ADC "includes filtering", but do not specify at all this filtering. Additionally reading some documentation of the ITU-T (ITU-T Software Tool Library Manual) I gathered that the G.711, G.721, and G.728 coders were developed using IRS weighting filtering (P.48) while modified IRS (P830 annex D) was used for the G.723.1 and the G.729. On the other hand, the specifications of G.729 says that "this coder is designed to operate with a digital signal obtained by first performing telephone bandwidth filtering (G.712) of the analogue input signal, then sampling it at 8 kHz". After reading all that, I gave a look to the ITU-T software tools library (G.191) and have found that I have, among others, the following options: 1- IRS weighting at 16 kHz, followed by high quality down-sampling by 2- Modified IRS weighting at 16 kHz, followed by high quality down-sampling by 2. 3- G.712 iir filter with down-sampling by 2. I am really confused and I do not know which option to select, to produce the input signal of the (GSM EFR) coder from my 16 kHz database. Additionally, one of the reviewer of a paper I am writing have asked: "it is suggested that the authors, add in the final paper, results using a filtered version of TIMIT8K in the telephony frequency range (300-3400 Hz)." For this request, I feel like using option 3. Sounds correct? Any help/comments would be appreciated, Thanks a lot, Sara |
Filtering for input signals to the GSM EFR coder
Started by ●April 23, 2000