CMU Sphinx / Forums / Help: Low performance with the official German model and a flac/wav audio sample.

I am trying to use pocketsphinx_continuous but I get bad results. I am using the official model. I created a GitHub repository with the exact models and cli-arguments I used and with the results I got. See: https://github.com/inktrap/cmusphinx-de

Update pocketsphinx & sphinxbase from github, results would be

INFO: cmn_live.c(120): Update from < 40.00  3.00 -1.00  0.00  0.00  0.00  0.00  0.00  0.00  0.00  0.00  0.00  0.00 >
INFO: cmn_live.c(138): Update to   < 45.11  8.09 -10.98 13.57 -0.91 -7.97  3.55 -7.29  0.96  1.32 -2.45 -4.99 -2.84 >
INFO: ngram_search_fwdtree.c(1550):     1102 words recognized (3/fr)
INFO: ngram_search_fwdtree.c(1552):   514495 senones evaluated (1234/fr)
INFO: ngram_search_fwdtree.c(1556):   632761 channels searched (1517/fr), 123146 1st, 30223 last
INFO: ngram_search_fwdtree.c(1559):     2287 words for which last channels evaluated (5/fr)
INFO: ngram_search_fwdtree.c(1561):    10760 candidate words for entering last phone (25/fr)
INFO: ngram_search_fwdtree.c(1564): fwdtree 0.63 CPU 0.152 xRT
INFO: ngram_search_fwdtree.c(1567): fwdtree 0.63 wall 0.152 xRT
INFO: ngram_search_fwdflat.c(302): Utterance vocabulary contains 47 words
INFO: ngram_search_fwdflat.c(948):      700 words recognized (2/fr)
INFO: ngram_search_fwdflat.c(950):    42290 senones evaluated (101/fr)
INFO: ngram_search_fwdflat.c(952):    26430 channels searched (63/fr)
INFO: ngram_search_fwdflat.c(954):     3013 words searched (7/fr)
INFO: ngram_search_fwdflat.c(957):     2376 word transitions (5/fr)
INFO: ngram_search_fwdflat.c(960): fwdflat 0.04 CPU 0.010 xRT
INFO: ngram_search_fwdflat.c(963): fwdflat 0.04 wall 0.010 xRT
INFO: ngram_search.c(1250): lattice start node <s>.0 end node </s>.412
INFO: ngram_search.c(1276): Eliminated 0 nodes before end node
INFO: ngram_search.c(1381): Lattice has 104 nodes, 97 links
INFO: ps_lattice.c(1380): Bestpath score: -8458
INFO: ps_lattice.c(1384): Normalizer P(O) = alpha(</s>:412:415) = -633741
INFO: ps_lattice.c(1441): Joint P(O,S) = -646082 P(S|O) = -12341
INFO: ngram_search.c(872): bestpath 0.00 CPU 0.000 xRT
INFO: ngram_search.c(875): bestpath 0.00 wall 0.000 xRT
eine gewisse temperatur ist notwendig

Make sure you get the same cmn_live.c line.

inktrap - 2017-01-30

Wow! That was fast! Thanks! (I used the debian version because stable is recommended in the documentation and they are the same).

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

INFO: cmn_live.c(120): Update from < 22.11 -11.62 -4.30 -2.35 -4.95 -0.05 -2.76  1.20 -2.92  2.77 -2.43  0.89 -1.09 >
INFO: cmn_live.c(138): Update to   < 46.72  8.88 -11.39 14.13 -0.57 -8.38  3.76 -7.83  0.98  1.18 -2.50 -5.38 -2.70 >
INFO: ngram_search_fwdtree.c(1550):     4209 words recognized (13/fr)
INFO: ngram_search_fwdtree.c(1552):   666189 senones evaluated (2044/fr)
INFO: ngram_search_fwdtree.c(1556):  1174628 channels searched (3603/fr), 111830 1st, 106531 last
INFO: ngram_search_fwdtree.c(1559):     6909 words for which last channels evaluated (21/fr)
INFO: ngram_search_fwdtree.c(1561):    47796 candidate words for entering last phone (146/fr)
INFO: ngram_search_fwdtree.c(1564): fwdtree 0.98 CPU 0.302 xRT
INFO: ngram_search_fwdtree.c(1567): fwdtree 0.99 wall 0.303 xRT
INFO: ngram_search_fwdflat.c(302): Utterance vocabulary contains 204 words
INFO: ngram_search_fwdflat.c(948):     3024 words recognized (9/fr)
INFO: ngram_search_fwdflat.c(950):   232229 senones evaluated (712/fr)
INFO: ngram_search_fwdflat.c(952):   236943 channels searched (726/fr)
INFO: ngram_search_fwdflat.c(954):    17093 words searched (52/fr)
INFO: ngram_search_fwdflat.c(957):    12251 word transitions (37/fr)
INFO: ngram_search_fwdflat.c(960): fwdflat 0.30 CPU 0.091 xRT
INFO: ngram_search_fwdflat.c(963): fwdflat 0.29 wall 0.090 xRT
INFO: ngram_search.c(1197): </s> not found in last frame, using eine.324 instead
INFO: ngram_search.c(1250): lattice start node <s>.0 end node eine.284
INFO: ngram_search.c(1276): Eliminated 41 nodes before end node
INFO: ngram_search.c(1381): Lattice has 270 nodes, 342 links
INFO: ps_lattice.c(1380): Bestpath score: -17548
INFO: ps_lattice.c(1384): Normalizer P(O) = alpha(eine:284:324) = -1197107
INFO: ps_lattice.c(1441): Joint P(O,S) = -1212093 P(S|O) = -14986
INFO: ngram_search.c(872): bestpath 0.00 CPU 0.000 xRT
INFO: ngram_search.c(875): bestpath 0.00 wall 0.000 xRT
nein wir sehr dankbar tue es notwendig eine
INFO: cmn_live.c(120): Update from < 46.72  8.88 -11.39 14.13 -0.57 -8.38  3.76 -7.83  0.98  1.18 -2.50 -5.38 -2.70 >
INFO: cmn_live.c(138): Update to   < 46.72  8.88 -11.39 14.13 -0.57 -8.38  3.76 -7.83  0.98  1.18 -2.50 -5.38 -2.70 >
INFO: ngram_search_fwdflat.c(302): Utterance vocabulary contains 0 words
INFO: ngram_search_fwdtree.c(429): TOTAL fwdtree 1.18 CPU 0.305 xRT
INFO: ngram_search_fwdtree.c(432): TOTAL fwdtree 1.18 wall 0.306 xRT
INFO: ngram_search_fwdflat.c(176): TOTAL fwdflat 0.30 CPU 0.076 xRT
INFO: ngram_search_fwdflat.c(179): TOTAL fwdflat 0.30 wall 0.076 xRT
INFO: ngram_search.c(303): TOTAL bestpath 0.00 CPU 0.001 xRT
INFO: ngram_search.c(306): TOTAL bestpath 0.00 wall 0.000 xRT

I installed the recent versions (User: cmusphinx)

inktrap - 2017-01-30

The cmn_live.c lines don't match.

Btw.: Here is the script:

#!/bin/bash #pocketsphinx_continuous -inmic yes -lm cmusphinx-voxforge-de.lm -dict cmusphinx-voxforge-de.dic -hmm cmusphinx-de-voxforge-5.2/ pocketsphinx_continuous -infile ../test.wav -lm cmusphinx-voxforge-de.lm.bin -dict cmusphinx-voxforge-de.dic -hmm cmusphinx-de-voxforge-5.2/

Last edit: inktrap 2017-01-30
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

If I am using cmusphinx-de-ptm-voxforge-5.2 I get

INFO: cmn_live.c(120): Update from < 22.11 -11.62 -4.30 -2.35 -4.95 -0.05 -2.76  1.20 -2.92  2.77 -2.43  0.89 -1.09 >
INFO: cmn_live.c(138): Update to   < 46.72  8.88 -11.39 14.13 -0.57 -8.38  3.76 -7.83  0.98  1.18 -2.50 -5.38 -2.70 >
INFO: ngram_search.c(467): Resized score stack to 200000 entries
INFO: ngram_search_fwdtree.c(1550):     4414 words recognized (14/fr)
INFO: ngram_search_fwdtree.c(1552):   478065 senones evaluated (1466/fr)
INFO: ngram_search_fwdtree.c(1556):   961092 channels searched (2948/fr), 88996 1st, 145539 last
INFO: ngram_search_fwdtree.c(1559):     7619 words for which last channels evaluated (23/fr)
INFO: ngram_search_fwdtree.c(1561):    37318 candidate words for entering last phone (114/fr)
INFO: ngram_search_fwdtree.c(1564): fwdtree 0.50 CPU 0.152 xRT
INFO: ngram_search_fwdtree.c(1567): fwdtree 0.50 wall 0.154 xRT
INFO: ngram_search_fwdflat.c(302): Utterance vocabulary contains 282 words
INFO: ngram_search_fwdflat.c(948):     2464 words recognized (8/fr)
INFO: ngram_search_fwdflat.c(950):   197682 senones evaluated (606/fr)
INFO: ngram_search_fwdflat.c(952):   287464 channels searched (881/fr)
INFO: ngram_search_fwdflat.c(954):    19121 words searched (58/fr)
INFO: ngram_search_fwdflat.c(957):    16076 word transitions (49/fr)
INFO: ngram_search_fwdflat.c(960): fwdflat 0.15 CPU 0.047 xRT
INFO: ngram_search_fwdflat.c(963): fwdflat 0.16 wall 0.048 xRT
INFO: ngram_search.c(1197): </s> not found in last frame, using notwendigen.324 instead
INFO: ngram_search.c(1250): lattice start node <s>.0 end node notwendigen.212
INFO: ngram_search.c(1276): Eliminated 114 nodes before end node
INFO: ngram_search.c(1381): Lattice has 350 nodes, 67 links
INFO: ps_lattice.c(1380): Bestpath score: -10204
INFO: ps_lattice.c(1384): Normalizer P(O) = alpha(notwendigen:212:324) = -935518
INFO: ps_lattice.c(1441): Joint P(O,S) = -943222 P(S|O) = -7704
INFO: ngram_search.c(872): bestpath 0.00 CPU 0.000 xRT
INFO: ngram_search.c(875): bestpath 0.00 wall 0.000 xRT
eine gewisse temperatur des notwendigen
INFO: cmn_live.c(120): Update from < 46.72  8.88 -11.39 14.13 -0.57 -8.38  3.76 -7.83  0.98  1.18 -2.50 -5.38 -2.70 >
INFO: cmn_live.c(138): Update to   < 46.72  8.88 -11.39 14.13 -0.57 -8.38  3.76 -7.83  0.98  1.18 -2.50 -5.38 -2.70 >
INFO: ngram_search_fwdflat.c(302): Utterance vocabulary contains 0 words
INFO: ngram_search_fwdtree.c(429): TOTAL fwdtree 0.60 CPU 0.155 xRT
INFO: ngram_search_fwdtree.c(432): TOTAL fwdtree 0.62 wall 0.160 xRT
INFO: ngram_search_fwdflat.c(176): TOTAL fwdflat 0.16 CPU 0.040 xRT
INFO: ngram_search_fwdflat.c(179): TOTAL fwdflat 0.16 wall 0.041 xRT
INFO: ngram_search.c(303): TOTAL bestpath 0.00 CPU 0.000 xRT
INFO: ngram_search.c(306): TOTAL bestpath 0.00 wall 0.000 xRT

Nickolay V. Shmyrev - 2017-02-01

You probably edited the model file feat.params somehow. It should start with 40,3,-1.

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

I tried it again with the current models from https://sourceforge.net/projects/cmusphinx/files/Acoustic%20and%20Language%20Models/German/ and get same output

eine gewisse temperatur des notwendigen

running:

#!/bin/bash

pocketsphinx_continuous -infile ../test.wav -lm cmusphinx-voxforge-de.lm.bin -dict cmusphinx-voxforge-de.dic -hmm cmusphinx-de-ptm-voxforge-5.2

And I did not change feat.params:

vh@box ~/src/sphinx/cmusphinx-de/German [master●●][i] % md5sum cmusphinx-de-ptm-voxforge-5.2/feat.params cmusphinx-de-ptm-voxforge-5.2.bak/feat.params 
275be8550f70128299c0320cba9db178  cmusphinx-de-ptm-voxforge-5.2/feat.params
275be8550f70128299c0320cba9db178  cmusphinx-de-ptm-voxforge-5.2.bak/feat.params

Last edit: inktrap 2017-02-09

ptm.log

inktrap - 2017-02-09

Could you share your working configuration?
Please also have a look at my repository here:
https://github.com/inktrap/cmusphinx-de

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
- Nickolay V. Shmyrev - 2017-02-10
  
  Add -vad_threshold 3.5 and it should work as expected.
  
  If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

Low performance with the official German model and a flac/wav audio sample.

Speech Recognition Toolkit

Forums

Help

Low performance with the official German model and a flac/wav audio sample. document.SUBSCRIPTION_OPTIONS = { "thing": "topic", "subscribed": false, "url": "subscribe", "icon": { "css": "fa fa-envelope-o" } };

Low performance with the official German model and a flac/wav audio sample.