![]() This works well on high-noise data, but tends to make voicing false-detections on more conventional low-noise data (such as telephone speech). The "rats" configs are trained on high-noise radio channel data (from the RATS project). The "keele" config is trained on the well-known Keele pitch database, with added pink noise between 0 and 20dB SNR. They differ in the sampling rate and structure of the subband filterbank (audio files are transparently resampled to the appropriate sampling rate, so this is a free choice), and in the data used to train the pitch classifier. The package above includes four different classifiers, corresponding to the four different config files referenced in the examples. The original Matlab code used to build this compiled target is available at Īll sources and the parameter files are in the package SAcC-v1.74.zip.įeel free to contact me with any problems. You'll still need to download the source package below to get the parameter files. You will also need to download and install the Matlab Compiler Runtime (MCR) Installer. This package has been compiled for several targets using the Matlab compiler. Train_mkMLP: qnstrn ftr1_file=keeleclean-sbac.pf hardtarget_file=keeleclean-ptch67.pf hardtarget_format=pfile ftr1_norm_file=keeleclean-sbac.norms ftr1_ftr_start=0 ftr1_ftr_count=240 window_extent=1 mlp3_input_size=240 mlp3_hidden_size=100 mlp3_output_size=68 train_sent_range=0:7 train_cache_frames=100000 cv_sent_range=8:9 learnrate_vals=0.008 ftr1_window_len=1 ftr1_window_offset=0 ftr1_delta_order=0 ftr1_delta_win=1 out_weight_file=keeleclean-sbac-ptch67-h100.wgt hardtarget_window_offset=0 log_file=keeleclean.log log_weight_file=keeleclean.chklog ckpt_weight_file=keeleclean.chk hardtarget_lastlab_reject=trueĬonfig params file written to keeleclean-config.txt Norms file written: keeleclean-sbac.norms Qnnorm was invoked with the following argument values Qnnorm norm_ftrfile=keeleclean-sbac.pf output_normfile=keeleclean-sbac.norms % We can now run the newly-trained pitch tracker, using the new % config file Train_SAcC(idlist, audiodir, audioext, gtdir, gtext, name) Idlist = textread(fullfile(pitchdir, 'idlist.txt'), '%s') These are part of the icsi-scenic-tools package. ![]() Note: to work, this routine requires working binaries for pfile_create, qnnorm, and qnstrn in the path. ![]() This routine also relies on the pt_read.m function from that package, so it needs to be installed in a sibling directory. We now provide functions to support the training of new SAcC classifiers from training sets consisting of audio and ground-truth pitch tracks, or possibly consensus pseudo-ground truth as produced by the pt_mkPseudoGt package. % Now load the pitch track file written by SAcC % (b) for files_b.list % afn = 'audio/BP_104.sph' % pfn = 'out/BP_' Load one of the example files and plot its pitch track % (a) for files.list SAcC files.list conf/rats_sr8k_bpo6_sb24_k10.config % (3) the faster new config sr=8k,bpo=6,sb=24,kdim=10 trained on Keele % SAcC files.list conf/keele_sr8k_bpo6_sb24_k10.config % (4) Babel example with Babelnet config sr=8k,bpo=6,sb=24,kdim=10 % SAcC files_b.list conf/Babelnet_sr8k_bpo6_sb24_k10.config % (5) Babel example with RATS config sr=8k,bpo=6,sb=24,kdim=10 % SAcC files_b.list conf/rats_sr8k_bpo6_sb24_k10.config % (6) Babel example with Keele config sr=8k,bpo=6,sb=24,kdim=10 % SAcC files_b.list conf/keele_sr8k_bpo6_sb24_k10.config Wrote ASCII-format out/ The Matlab script can be run from the Matlab prompt, or using the included Unix shell wrapper, run_SAcC.sh: Run it over our demo files % (1) the previous default sr=16k,bpo=16,sb=48,kdim=10 trained on RATS % SAcC files.list conf/rats_sr16k_bpo16_sb48_k10.config % (2) the faster new config sr=8k,bpo=6,sb=24,kdim=10 trained on RATS Load one of the example files and plot its pitch track.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |