I've seen the link : "From your keyword spotting results count how many false alarms and missed detections you’ve encountered.
Select the threshold with the smallest amount of false alarms and missed detections."
but i'm not sure about how to detect missed detection and false alarms from the result ...
For example : according to the result posted above if we count all the lines, there is 117 detection of "bonjour ubo" . And my file contains juste 17 occurence of this keyphrase
So with /1e-10/ we got so many false alarms whitch means this threshold is not good
- Is this how it should be done, or i'm missing somthing ?
- Should i choose smaller or bigger threshold, to reduce false alarms ?
Can you please explain more in details ?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi
I'm performing some tests using PocketSphinx keyword spotting tool with the following configuration (French langage) :
- acoustic model : fr-fr ;
- dictionary : fr.dict ;
- Keyword: "bonjour ubo" ;
- threshold: /1e-10/
- Audio file: keywords_record.wav ;
Using this command line :
pocketsphinx_continuous.exe -infile test\data\keywords_record.wav -keyphrase "bonjour ubo" -kws_threshold /1e-10/ -time yes -hmm model\fr-fr\fr-fr -dict model\fr-us\fr.dict
Gives me this ouput :
C:\sphinx\pocketsphinx>bin\Release\x64\pocketsphinx_continuous.exe -infile test\ data\keywords_record.wav -keyphrase "bonjour ubo" -kws_threshold \1e-10\ -time y
es -hmm model\fr-fr\fr-fr -dict model\fr-fr\fr.dict -logfn mbu_file.log
bonjour ubo
bonjour ubo 10.710 11.160 0.786921
bonjour ubo
bonjour ubo 12.590 12.890 0.764884
bonjour ubo
bonjour ubo 19.060 19.460 0.786685
bonjour ubo
bonjour ubo 66.630 67.040 0.789759
bonjour ubo
bonjour ubo 119.570 119.900 0.773423
bonjour ubo
bonjour ubo 199.390 199.910 0.796103
bonjour ubo
bonjour ubo 332.790 333.690 0.822898
bonjour ubo
bonjour ubo 337.690 338.090 0.791024
bonjour ubo
bonjour ubo 368.250 368.620 0.783230
bonjour ubo bonjour ubo
bonjour ubo 377.980 378.350 0.786606
bonjour ubo 374.490 375.010 0.778467
bonjour ubo bonjour ubo
bonjour ubo 391.940 392.250 0.759624
bonjour ubo 391.160 391.570 0.782917
bonjour ubo
bonjour ubo 466.980 467.260 0.773036
bonjour ubo
bonjour ubo 500.350 500.660 0.802739
bonjour ubo bonjour ubo
bonjour ubo 509.150 509.450 0.778856
bonjour ubo 508.780 509.020 0.765726
bonjour ubo
bonjour ubo 513.260 513.510 0.777300
bonjour ubo
bonjour ubo 517.880 518.130 0.763279
bonjour ubo
bonjour ubo 525.820 526.140 0.774894
bonjour ubo
bonjour ubo 539.500 539.790 0.785584
bonjour ubo
bonjour ubo 542.120 542.360 0.781431
bonjour ubo bonjour ubo bonjour ubo bonjour ubo
bonjour ubo 546.030 546.420 0.807005
bonjour ubo 545.490 545.840 0.782604
bonjour ubo 544.980 545.420 0.784171
bonjour ubo 544.440 544.680 0.775204
bonjour ubo
bonjour ubo 567.970 568.450 0.805876
bonjour ubo
bonjour ubo 639.120 640.000 0.810646
bonjour ubo
bonjour ubo 643.880 644.130 0.763814
bonjour ubo
bonjour ubo 656.810 657.070 0.787393
bonjour ubo
bonjour ubo 670.780 671.210 0.818548
bonjour ubo
bonjour ubo 684.500 684.790 0.788654
bonjour ubo
bonjour ubo 919.690 919.960 0.787629
bonjour ubo
bonjour ubo 923.990 924.330 0.787314
bonjour ubo bonjour ubo
bonjour ubo 931.020 931.370 0.786213
bonjour ubo 929.680 929.950 0.793163
bonjour ubo
bonjour ubo 963.330 963.680 0.771723
bonjour ubo
bonjour ubo 965.210 965.570 0.817403
bonjour ubo bonjour ubo bonjour ubo bonjour ubo bonjour ubo
bonjour ubo 973.160 973.410 0.777844
bonjour ubo 972.750 973.030 0.785663
bonjour ubo 972.160 972.470 0.787708
bonjour ubo 971.650 972.000 0.793242
bonjour ubo 971.150 971.520 0.809997
bonjour ubo
bonjour ubo 988.720 989.000 0.778389
bonjour ubo
bonjour ubo 998.260 998.550 0.759168
bonjour ubo
bonjour ubo 1014.240 1014.730 0.829260
bonjour ubo
bonjour ubo 1104.460 1104.750 0.766645
bonjour ubo bonjour ubo
bonjour ubo 1125.400 1125.710 0.759472
bonjour ubo 1124.240 1124.580 0.753043
bonjour ubo
bonjour ubo 1183.960 1184.280 0.775824
bonjour ubo
bonjour ubo 1221.050 1221.340 0.791578
bonjour ubo
bonjour ubo 1235.130 1235.860 0.786999
bonjour ubo
bonjour ubo 1244.040 1244.810 0.800735
bonjour ubo
bonjour ubo 1289.610 1289.880 0.767336
bonjour ubo
bonjour ubo 1304.540 1304.970 0.776989
bonjour ubo
bonjour ubo 1308.610 1309.570 0.813325
bonjour ubo
bonjour ubo 1336.290 1336.590 0.783779
bonjour ubo
bonjour ubo 1386.010 1386.330 0.790866
bonjour ubo
bonjour ubo 1392.300 1392.590 0.758409
bonjour ubo bonjour ubo
bonjour ubo 1439.500 1439.840 0.783074
bonjour ubo 1439.070 1439.380 0.783152
bonjour ubo
bonjour ubo 1443.040 1443.470 0.795785
bonjour ubo
bonjour ubo 1445.230 1445.720 0.809916
bonjour ubo
bonjour ubo 1449.740 1450.060 0.767182
bonjour ubo
bonjour ubo 1452.320 1452.620 0.786685
bonjour ubo
bonjour ubo 1457.080 1457.440 0.776833
bonjour ubo bonjour ubo
bonjour ubo 1459.170 1459.520 0.787078
bonjour ubo 1458.770 1459.020 0.766415
bonjour ubo
bonjour ubo 1498.560 1499.270 0.818957
bonjour ubo
bonjour ubo 1519.910 1520.200 0.784171
bonjour ubo
bonjour ubo 1641.490 1641.870 0.789917
bonjour ubo
bonjour ubo 1645.490 1645.820 0.810402
bonjour ubo
bonjour ubo 1647.750 1648.110 0.786370
bonjour ubo
bonjour ubo 1651.320 1651.680 0.807975
bonjour ubo
bonjour ubo 1865.540 1865.930 0.790470
bonjour ubo
bonjour ubo 1967.310 1967.640 0.774661
bonjour ubo
bonjour ubo 1983.270 1983.650 0.792687
bonjour ubo
bonjour ubo 1990.280 1990.800 0.816341
bonjour ubo
bonjour ubo 2049.750 2050.190 0.798016
bonjour ubo
bonjour ubo 2066.960 2067.220 0.772418
bonjour ubo
bonjour ubo 2162.340 2162.840 0.800014
bonjour ubo
bonjour ubo 2330.060 2330.500 0.789759
bonjour ubo
bonjour ubo 2345.410 2345.660 0.786291
bonjour ubo bonjour ubo
bonjour ubo 2677.070 2677.390 0.786370
bonjour ubo 2676.510 2676.870 0.796900
bonjour ubo 2690.870 2691.410 0.792608
bonjour ubo
bonjour ubo 2701.520 2701.840 0.775436
bonjour ubo
bonjour ubo 2717.570 2718.100 0.785977
bonjour ubo
bonjour ubo 2910.930 2911.450 0.781431
bonjour ubo
bonjour ubo 3049.110 3049.490 0.797538
bonjour ubo
bonjour ubo 3062.690 3063.120 0.792845
bonjour ubo
bonjour ubo 3108.040 3108.380 0.781978
bonjour ubo
bonjour ubo 3203.660 3203.950 0.773268
bonjour ubo
bonjour ubo 3215.530 3216.500 0.796183
bonjour ubo
bonjour ubo 3228.060 3228.480 0.779947
bonjour ubo
bonjour ubo 3241.630 3241.920 0.766186
bonjour ubo 3246.530 3247.050 0.778467
bonjour ubo
bonjour ubo 3251.300 3251.680 0.794910
bonjour ubo
bonjour ubo 3258.800 3259.160 0.771877
bonjour ubo
bonjour ubo 3262.680 3263.100 0.808864
bonjour ubo
bonjour ubo 3364.610 3364.930 0.779791
bonjour ubo
bonjour ubo 3485.530 3486.010 0.798655
bonjour ubo bonjour ubo
bonjour ubo 3517.290 3517.620 0.778233
bonjour ubo 3516.860 3517.200 0.799934
bonjour ubo
bonjour ubo 3520.850 3521.190 0.778389
bonjour ubo
bonjour ubo 3525.370 3525.840 0.788260
bonjour ubo
bonjour ubo 3529.320 3529.700 0.777455
bonjour ubo
bonjour ubo 3531.700 3532.210 0.791657
bonjour ubo
bonjour ubo 3539.000 3539.360 0.799774
bonjour ubo
bonjour ubo 3546.400 3546.740 0.778700
bonjour ubo
bonjour ubo 3550.980 3551.270 0.771568
bonjour ubo
bonjour ubo 3555.810 3556.070 0.772572
bonjour ubo
bonjour ubo 3560.780 3561.370 0.783152
bonjour ubo
bonjour ubo 3563.140 3563.600 0.807086
bonjour ubo
bonjour ubo 3567.720 3568.050 0.786213
bonjour ubo
bonjour ubo 3571.600 3571.960 0.776057
bonjour ubo
bonjour ubo 3578.680 3579.140 0.793718
So my questions are:
- what's the meaning of those numbers ?
- how can i use them to choose the best threshold ?
Thanks in advance :)
Last edit: mbu18 2017-10-26
Start time of detection, end time of detection and the confidence of detection.
Covered at https://cmusphinx.github.io/wiki/tutoriallm/#keyword-lists
Thanks for the reply !
I've seen the link :
"From your keyword spotting results count how many false alarms and missed detections you’ve encountered.
Select the threshold with the smallest amount of false alarms and missed detections."
but i'm not sure about how to detect missed detection and false alarms from the result ...
For example : according to the result posted above if we count all the lines, there is 117 detection of "bonjour ubo" . And my file contains juste 17 occurence of this keyphrase
So with /1e-10/ we got so many false alarms whitch means this threshold is not good
- Is this how it should be done, or i'm missing somthing ?
- Should i choose smaller or bigger threshold, to reduce false alarms ?
Can you please explain more in details ?
Yes
Try 1e-1