Filip Jurčíček

Free Internet TV, a 2008 link directory that outlived its era

2026-07-28T12:05:45.592+02:00

Another one from the archive. Free Internet TV is a directory of live TV channels streaming for free on the web, sorted by language and region — Arabic, Czech, English, French, German, Italian, Portuguese, Russian, Slovak.

It was built when watching television in a browser still meant installing RealPlayer and hoping your ADSL line held up. Plenty of the links have rotted since, but the surprise is how many still resolve, and how useful the list remains if you want news in a language you are learning rather than the algorithm's idea of what you should watch.

Like the Gmail Backup site, it now runs as a static mirror on S3 behind HTTPS. Old sites do not have to die; they just have to stop needing a database.

Gmail Backup is still online

2026-07-28T12:05:01.732+02:00

Jan Švec and I wrote Gmail Backup back in 2008 — a one-click tool that pulled your Gmail down over IMAP into a local folder, with incremental backups so you only fetched what was new. It ran on Windows and Linux, the code went out under GPL v3, and for a few years it was the answer people gave when someone asked how to get their mail off Google's servers.

It is old software and the world has moved on, but the site is still up: www.gmail-backup.com, with the downloads, the documentation, the FAQ and the forum threads intact. I recently moved the whole thing to static hosting and put it back on HTTPS, which is the only reason it still loads at all.

Worth remembering the premise, which has not aged: your mail is only yours if you have a copy of it.

Voice AI consulting through Voicebre Systems

2026-07-28T12:04:09.433+02:00

A note on what I do when I am not writing about KALDI graphs: I take on voice AI work through Voicebre Systems, a small consultancy based in Prague.

Three things, mostly: custom speech recognition — including voice commands and languages that the off-the-shelf stacks handle badly; natural language understanding for conversational systems; and end-to-end voice agents, from VUI design through to integration with whatever is already running in the business.

The interesting problems are usually not the model. They are the data you are allowed to use, the latency budget, and what happens on the turn where the user says something nobody anticipated. If that is the shape of your problem, get in touch.

Vokarys: dictation that runs on your own machine

2026-07-28T12:03:16.459+02:00

Most dictation today is a round trip to someone else's GPU. That is fine until your voice is the part the model was never trained on — an accent, a speech difference, a bad microphone in a noisy room — and you have no way to make it adapt.

That is the problem Vokarys goes after. It is a desktop app that turns speech into text with recognition running locally on your computer, with punctuation and capitalisation added as you talk. It learns from your own voice over time, and recordings stay on the machine unless you explicitly choose to contribute them. There is a voice chat mode wired to Gemini as well, and a browser version if you would rather not install anything.

If you have given up on dictation because the generic model never quite got you, it is worth a look.

Privacy policy

2024-02-15T19:35:00.008+01:00

Effective date: 2024-02-16

This Privacy Policy applies to the website [bozskyfilip.blogspot.com] operated by [Filip Jurcicek] ("us", "we", or "our"). This Policy informs you of our policies regarding the collection, use, and disclosure of personal data we receive from users of our website.

1. Information Collection and Use

We may collect and process the following types of data:

1.1 Personal Data:

We may collect personally identifiable information including, but not limited to, your name, email address, and contact information. We collect this information for the purpose of providing and improving our services to you.

1.2 Usage Data:

We may collect non-personally identifiable information about how you use our website. This may include your IP address, browser type, device type, operating system, and other similar information. We collect this information to analyze trends, administer the site, track users' movements, and gather demographic information for aggregate use.

2. Use of Data

We use your personal data for the following purposes:

2.1 To provide and maintain the website,

2.2 To notify you about changes to our website,

2.3 To allow you to participate in interactive features of our website when you choose to do so,

2.4 To provide customer support,

2.5 To gather analysis or valuable information so that we can improve our website,

2.6 To monitor the usage of our website,

2.7 To detect, prevent and address technical issues,

3. Disclosure of Data

We may disclose personal data in the good faith belief that such action is necessary to:

3.1 Comply with a legal obligation,

3.2 Protect and defend our rights or property,

3.3 Prevent or investigate possible wrongdoing,

3.4 Protect the personal safety of users of the website or the public,

3.5 Protect against legal liability.

We may also share non-personally identifiable information with third-party service providers to help us analyze how our website is used and improve our services. These third-party service providers are obligated to keep your information confidential.

4. Third-Party Services

This website uses Google AdSense AdWords to serve advertisements. Please note that Google AdSense AdWords uses cookies to provide personalized ads based on your interests. You can opt-out of Google AdSense AdWords services by visiting the Google Ads Settings page.

5. Security

We take reasonable measures to protect your personal data; however, no method of transmission over the internet or electronic storage is completely secure. We cannot guarantee the absolute security of your data.

6. Your Rights

You have the right to:

6.1 Request access and receive a copy of the personal data we hold about you,

6.2 Request rectification of any inaccurate personal data we hold about you,

6.3 Request erasure of your personal data,

6.4 Object to the processing of your personal data,

6.5 Request the restriction of processing of your personal data,

6.6 Request the transfer of your personal data to another party.

If you wish to exercise any of these rights, please contact us using the contact information provided at the end of this Privacy Policy.

7. Children's Privacy

Our website is not intended for use by individuals under the age of 13. We do not knowingly collect personally identifiable information from children under 13. If you are a parent or guardian and you areaware that your child has provided us with personal data, please contact us so that we can take necessary action to delete such information from our servers.

8. Changes to This Privacy Policy

We may update our Privacy Policy from time to time. We will notify you of any changes by posting the new Privacy Policy on this page. It is advised to review this Privacy Policy periodically for any changes. Changes are effective immediately upon posting on this page.

9. Contact Us

If you have any questions or concerns regarding this Privacy Policy, please contact us at [filip.jurcicek@gmail.com].

Please note that this privacy policy is designed to comply with the General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA). This policy outlines the information we collect, how we use it, and offers transparency on our usage of Google AdSense AdWords. By using our website, you agree to the terms outlined in this Privacy Policy.

Last updated: 2024-02-15

Categorical Reparameterization with Gumbel-Softmax & The Concrete Distri...

2020-07-02T17:18:00.002+02:00

Midnight Commander shortcuts

2018-02-20T09:38:00.000+01:00

Key	Action	Notes
Ctrl+o	toggle panes on/off
Ctrl+l	redraw screen	This is on all terminals
Ctrl+PgUp	goto parent dir
Ctrl+Enter	copy selected filename to command line	%f is equivalent
Ctrl+x+p	copy unselected panel's path to command line	%D is equivalent
Ctrl+x !	External panelize	Display paths returned from external command
Shift+mouse	select text
Insert	toggle selection of highlighted file
*	toggle selection
+	add pattern to selection
-	remove pattern from selection
F3	view
F4	edit
F5	copy
F6	rename
F7	mkdir
F8	remove
F9	menu
F10	Exit

how the make HCL and G graphs, and on the fly compositon of HCL and G for KALDI

2017-11-14T09:29:00.000+01:00

Well, I had again to do something ;-) The task is to generate/create/update a decoding graph for KALDI on the fly. In my case, I aim at changing a G (grammar) in the context of a dialogue system.

One can generate a new HCLG but this would take a lot of time as this involves FST determinization, epsilon-removal, minimization, etc. Therefore, I tried to use on-the-fly composition of statically prepared HCL and G. At first, I struggled with it but later I made it work. See https://github.com/jpuigcerver/kaldi-decoders/issues/1

Here is a short summary:

At the end, I managed to get LabelLookAheadMatcher to work. It is mostly based on the code and examples in opendcd, e.g. https://github.com/opendcd/opendcd/blob/master/script/makegraphotf.sh.

First, Here is how I build and prepare the HCL and G. Please not that OpenFST must be compiled with --enable-lookahead-fsts, see http://www.openfst.org/twiki/bin/view/FST/ReadMe.

#---------------

fstdeterminize ${lang}/L_disambig.fst | fstarcsort > ${dir}/det.L.fst

#---------------

fstcomposecontext \
    --context-size=$N --central-position=$P \
    --read-disambig-syms=${lang}/phones/disambig.int \
    --write-disambig-syms=${lang}/disambig_ilabels_${N}_${P}.int \
    ${dir}/ilabels_${N}_${P} ${dir}/det.L.fst | \
    fstarcsort > ${dir}/CL.fst

#---------------

make-h-transducer \
    --disambig-syms-out=${dir}/h.disambig.int \
    --transition-scale=$tscale \
    ${dir}/ilabels_${N}_${P} \
    ${tree} \
    ${model} > ${dir}/Ha.fst

cat ${dir}/Ha.fst > ${dir}/det.Ha.fst

#---------------

fstconvert \
     --fst_type=ilabel_lookahead \
     --save_relabel_ipairs=${dir}/h.orelabel ${dir}/CL.fst |
     fstarcsort --sort_type=ilabel > ${dir}/la.CL.fst
    
fstrelabel --relabel_opairs=${dir}/h.orelabel ${dir}/det.Ha.fst | \
     fstarcsort --sort_type=olabel | \
     fstcompose - ${dir}/la.CL.fst > ${dir}/det.HaCL.fst

#---------------

fstdeterminize ${dir}/det.HaCL.fst | \
    fstrmsymbols ${dir}/h.disambig.int | \
    fstrmepslocal | \
    fstpushspecial | \
    fstminimizeencoded | \
    add-self-loops --self-loop-scale=$loopscale --reorder=true ${model} - | 
    fstarcsort --sort_type=olabel |
    fstconvert --fst_type=const > ${dir}/HCL.fst

#-----------------------------

fstconvert --fst_type=olabel_lookahead --save_relabel_opairs=${dir}/g.irelabel ${dir}/HCL.fst > ${dir}/HCLr.fst
fstrelabel --relabel_ipairs=${dir}/g.irelabel ${lang}/G.fst | \
    fstarcsort | 
    fstconvert --fst_type=const > ${dir}/Gr.fst

fstcompose ${dir}/HCLr.fst ${dir}/Gr.fst | \
    fstconvert --fst_type=const > ${dir}/HCLrGr.fst

Please note that the HCLrGr.fst FST is here for testing purposes just to determine if offline composed HCLG is good. This can be simply tested using the regular code/decoder not assuming on-the-fly composition.

The composed HCLG for the KALDI decoder is created as follows:

ComposeFst* OTFComposeFst(
    const StdFst &ifst1, const StdFst &ifst2,
    const CacheOptions& cache_opts = CacheOptions()) {

  typedef LookAheadMatcher< StdFst > M;
  typedef AltSequenceComposeFilter SF;
  typedef LookAheadComposeFilter  LF;
  typedef PushWeightsComposeFilter WF;
  typedef PushLabelsComposeFilter ComposeFilter;
  typedef M FstMatcher;
  
  ComposeFstOptions opts(cache_opts);

  return new ComposeFst(ifst1, ifst2, opts);
}

My observation is that when I want the same WER then I must lower pruning for the OTF composed HCL and G. This results in about 20 % increase in RTF. If I fix RTF then my WER is about 20 % relatively worse for OTF composed HCL and G. So, there is some cost of the OTF composition though it is not that bad. It is usable.

Please note that preparation of the HCL and G is a bit different from the one in https://github.com/opendcd/opendcd/blob/master/script/makegraphotf.sh . For example, I could not determinize Ha.fst as it appeared to be non-functional. Also, the determinization of L is important, otherwise the final HCL graph will not be "small enough" and therefore the OTF composition would not be that efficient.

kaldi editing nnet3 chain model - adding a softmax layer on top of the chain output

2017-10-18T15:10:00.004+02:00

I had to do one more thing: to edit a trained kaldi nnet3 chain model and add a softmax layer on top of the chain model. The reason for this is to get "probability" like output directly from the chain model

First, let's look at the nnet structure:

nnet3-am-info final.mdl
input-dim: 20
ivector-dim: -1
num-pdfs: 6105
prior-dimension: 0
# Nnet info follows.
left-context: 15
right-context: 15
num-parameters: 15499085
modulus: 1
input-node name=input dim=20
component-node name=L0_fixaffine component=L0_fixaffine input=Append(Offset(input, -1), input, Offset(input, 1)) input-dim=60 output-dim=60
component-node name=Tdnn_0_affine component=Tdnn_0_affine input=L0_fixaffine input-dim=60 output-dim=625
component-node name=Tdnn_0_relu component=Tdnn_0_relu input=Tdnn_0_affine input-dim=625 output-dim=625
component-node name=Tdnn_0_renorm component=Tdnn_0_renorm input=Tdnn_0_relu input-dim=625 output-dim=625
component-node name=Tdnn_1_affine component=Tdnn_1_affine input=Append(Offset(Tdnn_0_renorm, -1), Tdnn_0_renorm, Offset(Tdnn_0_renorm, 1)) input-dim=1875 output-dim=625
component-node name=Tdnn_1_relu component=Tdnn_1_relu input=Tdnn_1_affine input-dim=625 output-dim=625
component-node name=Tdnn_1_renorm component=Tdnn_1_renorm input=Tdnn_1_relu input-dim=625 output-dim=625
component-node name=Tdnn_2_affine component=Tdnn_2_affine input=Append(Offset(Tdnn_1_renorm, -1), Tdnn_1_renorm, Offset(Tdnn_1_renorm, 1)) input-dim=1875 output-dim=625
component-node name=Tdnn_2_relu component=Tdnn_2_relu input=Tdnn_2_affine input-dim=625 output-dim=625
component-node name=Tdnn_2_renorm component=Tdnn_2_renorm input=Tdnn_2_relu input-dim=625 output-dim=625
component-node name=Tdnn_3_affine component=Tdnn_3_affine input=Append(Offset(Tdnn_2_renorm, -3), Tdnn_2_renorm, Offset(Tdnn_2_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_3_relu component=Tdnn_3_relu input=Tdnn_3_affine input-dim=625 output-dim=625
component-node name=Tdnn_3_renorm component=Tdnn_3_renorm input=Tdnn_3_relu input-dim=625 output-dim=625
component-node name=Tdnn_4_affine component=Tdnn_4_affine input=Append(Offset(Tdnn_3_renorm, -3), Tdnn_3_renorm, Offset(Tdnn_3_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_4_relu component=Tdnn_4_relu input=Tdnn_4_affine input-dim=625 output-dim=625
component-node name=Tdnn_4_renorm component=Tdnn_4_renorm input=Tdnn_4_relu input-dim=625 output-dim=625
component-node name=Tdnn_5_affine component=Tdnn_5_affine input=Append(Offset(Tdnn_4_renorm, -3), Tdnn_4_renorm, Offset(Tdnn_4_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_5_relu component=Tdnn_5_relu input=Tdnn_5_affine input-dim=625 output-dim=625
component-node name=Tdnn_5_renorm component=Tdnn_5_renorm input=Tdnn_5_relu input-dim=625 output-dim=625
component-node name=Tdnn_6_affine component=Tdnn_6_affine input=Append(Offset(Tdnn_5_renorm, -3), Tdnn_5_renorm, Offset(Tdnn_5_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_6_relu component=Tdnn_6_relu input=Tdnn_6_affine input-dim=625 output-dim=625
component-node name=Tdnn_6_renorm component=Tdnn_6_renorm input=Tdnn_6_relu input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_chain_affine component=Tdnn_pre_final_chain_affine input=Tdnn_6_renorm input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_chain_relu component=Tdnn_pre_final_chain_relu input=Tdnn_pre_final_chain_affine input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_chain_renorm component=Tdnn_pre_final_chain_renorm input=Tdnn_pre_final_chain_relu input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_xent_affine component=Tdnn_pre_final_xent_affine input=Tdnn_6_renorm input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_xent_relu component=Tdnn_pre_final_xent_relu input=Tdnn_pre_final_xent_affine input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_xent_renorm component=Tdnn_pre_final_xent_renorm input=Tdnn_pre_final_xent_relu input-dim=625 output-dim=625
component-node name=Final_affine component=Final_affine input=Tdnn_pre_final_chain_renorm input-dim=625 output-dim=6105
output-node name=output input=Final_affine dim=6105 objective=linear
component-node name=Final-xent_affine component=Final-xent_affine input=Tdnn_pre_final_xent_renorm input-dim=625 output-dim=6105
component-node name=Final-xent_log_softmax component=Final-xent_log_softmax input=Final-xent_affine input-dim=6105 output-dim=6105
output-node name=output-xent input=Final-xent_log_softmax dim=6105 objective=linear

...

The solution is a bit more complicated then the solution for editing task in the previous blog post.
First, I have to add some components to the network to add softmax layer, then to remove and rename the added softmax layer. This done trough creating nnet-config and then using edits provided by nnet3-am-copy:

echo "component-node name=Final_log_softmax component=Final_log_softmax input=Final_affine
output-node name=output-softmax input=Final_log_softmax objective=linear
component name=Final_log_softmax type=LogSoftmaxComponent dim=6105" > add.chain.softmax.cfg

cat final.mdl | nnet3-am-copy --nnet-config=add.chain.softmax.cfg - - | nnet3-am-copy --edits='remove-output-nodes name=output;rename-node old-name=output-softmax new-name=output;remove-orphans' - final.chain.softmax.mdl

LOG (nnet3-am-copy:main():nnet3-am-copy.cc:156) Copied neural net from - to -
LOG (nnet3-am-copy:ReadEditConfig():nnet-utils.cc:687) Removing 1 output nodes.
LOG (nnet3-am-copy:RemoveSomeNodes():nnet-nnet.cc:885) Removed 1 orphan nodes.
LOG (nnet3-am-copy:RemoveOrphanComponents():nnet-nnet.cc:810) Removing 0 orphan components.
LOG (nnet3-am-copy:main():nnet3-am-copy.cc:156) Copied neural net from - to final.chain.softmax.mdl

Finally, let's look at the modified structure:

nnet3-am-info final.chain.softmax.mdl
input-dim: 20
ivector-dim: -1
num-pdfs: 6105
prior-dimension: 0
# Nnet info follows.
left-context: 15
right-context: 15
num-parameters: 15499085
modulus: 1
input-node name=input dim=20
component-node name=L0_fixaffine component=L0_fixaffine input=Append(Offset(input, -1), input, Offset(input, 1)) input-dim=60 output-dim=60
component-node name=Tdnn_0_affine component=Tdnn_0_affine input=L0_fixaffine input-dim=60 output-dim=625
component-node name=Tdnn_0_relu component=Tdnn_0_relu input=Tdnn_0_affine input-dim=625 output-dim=625
component-node name=Tdnn_0_renorm component=Tdnn_0_renorm input=Tdnn_0_relu input-dim=625 output-dim=625
component-node name=Tdnn_1_affine component=Tdnn_1_affine input=Append(Offset(Tdnn_0_renorm, -1), Tdnn_0_renorm, Offset(Tdnn_0_renorm, 1)) input-dim=1875 output-dim=625
component-node name=Tdnn_1_relu component=Tdnn_1_relu input=Tdnn_1_affine input-dim=625 output-dim=625
component-node name=Tdnn_1_renorm component=Tdnn_1_renorm input=Tdnn_1_relu input-dim=625 output-dim=625
component-node name=Tdnn_2_affine component=Tdnn_2_affine input=Append(Offset(Tdnn_1_renorm, -1), Tdnn_1_renorm, Offset(Tdnn_1_renorm, 1)) input-dim=1875 output-dim=625
component-node name=Tdnn_2_relu component=Tdnn_2_relu input=Tdnn_2_affine input-dim=625 output-dim=625
component-node name=Tdnn_2_renorm component=Tdnn_2_renorm input=Tdnn_2_relu input-dim=625 output-dim=625
component-node name=Tdnn_3_affine component=Tdnn_3_affine input=Append(Offset(Tdnn_2_renorm, -3), Tdnn_2_renorm, Offset(Tdnn_2_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_3_relu component=Tdnn_3_relu input=Tdnn_3_affine input-dim=625 output-dim=625
component-node name=Tdnn_3_renorm component=Tdnn_3_renorm input=Tdnn_3_relu input-dim=625 output-dim=625
component-node name=Tdnn_4_affine component=Tdnn_4_affine input=Append(Offset(Tdnn_3_renorm, -3), Tdnn_3_renorm, Offset(Tdnn_3_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_4_relu component=Tdnn_4_relu input=Tdnn_4_affine input-dim=625 output-dim=625
component-node name=Tdnn_4_renorm component=Tdnn_4_renorm input=Tdnn_4_relu input-dim=625 output-dim=625
component-node name=Tdnn_5_affine component=Tdnn_5_affine input=Append(Offset(Tdnn_4_renorm, -3), Tdnn_4_renorm, Offset(Tdnn_4_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_5_relu component=Tdnn_5_relu input=Tdnn_5_affine input-dim=625 output-dim=625
component-node name=Tdnn_5_renorm component=Tdnn_5_renorm input=Tdnn_5_relu input-dim=625 output-dim=625
component-node name=Tdnn_6_affine component=Tdnn_6_affine input=Append(Offset(Tdnn_5_renorm, -3), Tdnn_5_renorm, Offset(Tdnn_5_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_6_relu component=Tdnn_6_relu input=Tdnn_6_affine input-dim=625 output-dim=625
component-node name=Tdnn_6_renorm component=Tdnn_6_renorm input=Tdnn_6_relu input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_chain_affine component=Tdnn_pre_final_chain_affine input=Tdnn_6_renorm input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_chain_relu component=Tdnn_pre_final_chain_relu input=Tdnn_pre_final_chain_affine input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_chain_renorm component=Tdnn_pre_final_chain_renorm input=Tdnn_pre_final_chain_relu input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_xent_affine component=Tdnn_pre_final_xent_affine input=Tdnn_6_renorm input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_xent_relu component=Tdnn_pre_final_xent_relu input=Tdnn_pre_final_xent_affine input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_xent_renorm component=Tdnn_pre_final_xent_renorm input=Tdnn_pre_final_xent_relu input-dim=625 output-dim=625
component-node name=Final_affine component=Final_affine input=Tdnn_pre_final_chain_renorm input-dim=625 output-dim=6105
component-node name=Final-xent_affine component=Final-xent_affine input=Tdnn_pre_final_xent_renorm input-dim=625 output-dim=6105
component-node name=Final-xent_log_softmax component=Final-xent_log_softmax input=Final-xent_affine input-dim=6105 output-dim=6105
output-node name=output-xent input=Final-xent_log_softmax dim=6105 objective=linear
component-node name=Final_log_softmax component=Final_log_softmax input=Final_affine input-dim=6105 output-dim=6105
output-node name=output input=Final_log_softmax dim=6105 objective=linear
...

component name=Final_log_softmax type=LogSoftmaxComponent, dim=6105
...

Cool, it worked!

kaldi editing nnet3 chain model - using the auxiliary xent output as the main output

2017-10-18T11:44:00.000+02:00

I had a task to edit a trained kaldi nnet3 chain model so that the output node is the output-xent instead the original output.

First, let's look at the nnet structure:

nnet3-am-info final.mdl
input-dim: 20
ivector-dim: -1
num-pdfs: 6105
prior-dimension: 0
# Nnet info follows.
left-context: 15
right-context: 15
num-parameters: 15499085
modulus: 1
input-node name=input dim=20
component-node name=L0_fixaffine component=L0_fixaffine input=Append(Offset(input, -1), input, Offset(input, 1)) input-dim=60 output-dim=60
component-node name=Tdnn_0_affine component=Tdnn_0_affine input=L0_fixaffine input-dim=60 output-dim=625
component-node name=Tdnn_0_relu component=Tdnn_0_relu input=Tdnn_0_affine input-dim=625 output-dim=625
component-node name=Tdnn_0_renorm component=Tdnn_0_renorm input=Tdnn_0_relu input-dim=625 output-dim=625
component-node name=Tdnn_1_affine component=Tdnn_1_affine input=Append(Offset(Tdnn_0_renorm, -1), Tdnn_0_renorm, Offset(Tdnn_0_renorm, 1)) input-dim=1875 output-dim=625
component-node name=Tdnn_1_relu component=Tdnn_1_relu input=Tdnn_1_affine input-dim=625 output-dim=625
component-node name=Tdnn_1_renorm component=Tdnn_1_renorm input=Tdnn_1_relu input-dim=625 output-dim=625
component-node name=Tdnn_2_affine component=Tdnn_2_affine input=Append(Offset(Tdnn_1_renorm, -1), Tdnn_1_renorm, Offset(Tdnn_1_renorm, 1)) input-dim=1875 output-dim=625
component-node name=Tdnn_2_relu component=Tdnn_2_relu input=Tdnn_2_affine input-dim=625 output-dim=625
component-node name=Tdnn_2_renorm component=Tdnn_2_renorm input=Tdnn_2_relu input-dim=625 output-dim=625
component-node name=Tdnn_3_affine component=Tdnn_3_affine input=Append(Offset(Tdnn_2_renorm, -3), Tdnn_2_renorm, Offset(Tdnn_2_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_3_relu component=Tdnn_3_relu input=Tdnn_3_affine input-dim=625 output-dim=625
component-node name=Tdnn_3_renorm component=Tdnn_3_renorm input=Tdnn_3_relu input-dim=625 output-dim=625
component-node name=Tdnn_4_affine component=Tdnn_4_affine input=Append(Offset(Tdnn_3_renorm, -3), Tdnn_3_renorm, Offset(Tdnn_3_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_4_relu component=Tdnn_4_relu input=Tdnn_4_affine input-dim=625 output-dim=625
component-node name=Tdnn_4_renorm component=Tdnn_4_renorm input=Tdnn_4_relu input-dim=625 output-dim=625
component-node name=Tdnn_5_affine component=Tdnn_5_affine input=Append(Offset(Tdnn_4_renorm, -3), Tdnn_4_renorm, Offset(Tdnn_4_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_5_relu component=Tdnn_5_relu input=Tdnn_5_affine input-dim=625 output-dim=625
component-node name=Tdnn_5_renorm component=Tdnn_5_renorm input=Tdnn_5_relu input-dim=625 output-dim=625
component-node name=Tdnn_6_affine component=Tdnn_6_affine input=Append(Offset(Tdnn_5_renorm, -3), Tdnn_5_renorm, Offset(Tdnn_5_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_6_relu component=Tdnn_6_relu input=Tdnn_6_affine input-dim=625 output-dim=625
component-node name=Tdnn_6_renorm component=Tdnn_6_renorm input=Tdnn_6_relu input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_chain_affine component=Tdnn_pre_final_chain_affine input=Tdnn_6_renorm input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_chain_relu component=Tdnn_pre_final_chain_relu input=Tdnn_pre_final_chain_affine input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_chain_renorm component=Tdnn_pre_final_chain_renorm input=Tdnn_pre_final_chain_relu input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_xent_affine component=Tdnn_pre_final_xent_affine input=Tdnn_6_renorm input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_xent_relu component=Tdnn_pre_final_xent_relu input=Tdnn_pre_final_xent_affine input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_xent_renorm component=Tdnn_pre_final_xent_renorm input=Tdnn_pre_final_xent_relu input-dim=625 output-dim=625
component-node name=Final_affine component=Final_affine input=Tdnn_pre_final_chain_renorm input-dim=625 output-dim=6105
output-node name=output input=Final_affine dim=6105 objective=linear
component-node name=Final-xent_affine component=Final-xent_affine input=Tdnn_pre_final_xent_renorm input-dim=625 output-dim=6105
component-node name=Final-xent_log_softmax component=Final-xent_log_softmax input=Final-xent_affine input-dim=6105 output-dim=6105
output-node name=output-xent input=Final-xent_log_softmax dim=6105 objective=linear

...

The solution removes chain output node, removes orphans as some of the nodes are not necessary anymore, and finally renames the xent output. Now we can do the edits:

nnet3-am-copy --edits='remove-output-nodes name=output;remove-orphans;rename-node old-name=output-xent new-name=output' final.mdl final.xent.mdl

LOG (nnet3-am-copy:ReadEditConfig():nnet-utils.cc:687) Removing 1 output nodes.
LOG (nnet3-am-copy:RemoveSomeNodes():nnet-nnet.cc:885) Removed 1 orphan nodes.
LOG (nnet3-am-copy:RemoveSomeNodes():nnet-nnet.cc:885) Removed 8 orphan nodes.
LOG (nnet3-am-copy:RemoveOrphanComponents():nnet-nnet.cc:810) Removing 4 orphan components.
LOG (nnet3-am-copy:main():nnet3-am-copy.cc:156) Copied neural net from final.mdl to final.xent.mdl

Finally, let's look at the modified structure:

nnet3-am-info final.xent.mdl
input-dim: 20
ivector-dim: -1
num-pdfs: 6105
prior-dimension: 0
# Nnet info follows.
left-context: 15
right-context: 15
num-parameters: 11286105
modulus: 1
input-node name=input dim=20
component-node name=L0_fixaffine component=L0_fixaffine input=Append(Offset(input, -1), input, Offset(input, 1)) input-dim=60 output-dim=60
component-node name=Tdnn_0_affine component=Tdnn_0_affine input=L0_fixaffine input-dim=60 output-dim=625
component-node name=Tdnn_0_relu component=Tdnn_0_relu input=Tdnn_0_affine input-dim=625 output-dim=625
component-node name=Tdnn_0_renorm component=Tdnn_0_renorm input=Tdnn_0_relu input-dim=625 output-dim=625
component-node name=Tdnn_1_affine component=Tdnn_1_affine input=Append(Offset(Tdnn_0_renorm, -1), Tdnn_0_renorm, Offset(Tdnn_0_renorm, 1)) input-dim=1875 output-dim=625
component-node name=Tdnn_1_relu component=Tdnn_1_relu input=Tdnn_1_affine input-dim=625 output-dim=625
component-node name=Tdnn_1_renorm component=Tdnn_1_renorm input=Tdnn_1_relu input-dim=625 output-dim=625
component-node name=Tdnn_2_affine component=Tdnn_2_affine input=Append(Offset(Tdnn_1_renorm, -1), Tdnn_1_renorm, Offset(Tdnn_1_renorm, 1)) input-dim=1875 output-dim=625
component-node name=Tdnn_2_relu component=Tdnn_2_relu input=Tdnn_2_affine input-dim=625 output-dim=625
component-node name=Tdnn_2_renorm component=Tdnn_2_renorm input=Tdnn_2_relu input-dim=625 output-dim=625
component-node name=Tdnn_3_affine component=Tdnn_3_affine input=Append(Offset(Tdnn_2_renorm, -3), Tdnn_2_renorm, Offset(Tdnn_2_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_3_relu component=Tdnn_3_relu input=Tdnn_3_affine input-dim=625 output-dim=625
component-node name=Tdnn_3_renorm component=Tdnn_3_renorm input=Tdnn_3_relu input-dim=625 output-dim=625
component-node name=Tdnn_4_affine component=Tdnn_4_affine input=Append(Offset(Tdnn_3_renorm, -3), Tdnn_3_renorm, Offset(Tdnn_3_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_4_relu component=Tdnn_4_relu input=Tdnn_4_affine input-dim=625 output-dim=625
component-node name=Tdnn_4_renorm component=Tdnn_4_renorm input=Tdnn_4_relu input-dim=625 output-dim=625
component-node name=Tdnn_5_affine component=Tdnn_5_affine input=Append(Offset(Tdnn_4_renorm, -3), Tdnn_4_renorm, Offset(Tdnn_4_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_5_relu component=Tdnn_5_relu input=Tdnn_5_affine input-dim=625 output-dim=625
component-node name=Tdnn_5_renorm component=Tdnn_5_renorm input=Tdnn_5_relu input-dim=625 output-dim=625
component-node name=Tdnn_6_affine component=Tdnn_6_affine input=Append(Offset(Tdnn_5_renorm, -3), Tdnn_5_renorm, Offset(Tdnn_5_renorm, 3)) input-dim=1875 output-dim=625
component-node name=Tdnn_6_relu component=Tdnn_6_relu input=Tdnn_6_affine input-dim=625 output-dim=625
component-node name=Tdnn_6_renorm component=Tdnn_6_renorm input=Tdnn_6_relu input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_xent_affine component=Tdnn_pre_final_xent_affine input=Tdnn_6_renorm input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_xent_relu component=Tdnn_pre_final_xent_relu input=Tdnn_pre_final_xent_affine input-dim=625 output-dim=625
component-node name=Tdnn_pre_final_xent_renorm component=Tdnn_pre_final_xent_renorm input=Tdnn_pre_final_xent_relu input-dim=625 output-dim=625
component-node name=Final-xent_affine component=Final-xent_affine input=Tdnn_pre_final_xent_renorm input-dim=625 output-dim=6105
component-node name=Final-xent_log_softmax component=Final-xent_log_softmax input=Final-xent_affine input-dim=6105 output-dim=6105
output-node name=output input=Final-xent_log_softmax dim=6105 objective=linear
...

Cool, it worked!

Filip Jurčíček: „Zajímá mě přirozenost komunikace pomocí hlasu“

2015-01-21T11:37:00.001+01:00

Filip Jurčíček: „Zajímá mě přirozenost komunikace pomocí hlasu“:

'via Blog this'

Online SVM

2014-03-16T14:02:00.001+01:00

An exact solution to the problem of online SVM learning has been found by Cauwenberghs and Poggio (2001). Their incremental algorithm (hereinafter referred to as a C&P algorithm) updates an optimal solution of an SVM training problem after one training example is added (or removed).

'via Blog this'

How to maintain a successful open source project — Medium

2013-07-24T00:50:00.001+02:00

Very inspiring.

How to maintain a successful open source project — Medium:

'via Blog this'

Backup your hard drive over the network with dd and ssh

2013-06-19T20:53:00.001+02:00

Backup your hard drive over the network with dd and ssh:

Citation:

... wanted to copy everything off the disk and send it over the network. So we can do it with ssh. First zero out the non used space on the running disk to make compressing the image much eaiser. Using the command:

dd if=/dev/zero of=0bits bs=20M; rm 0bits

Then boot knoppix (or any other bootable linux distro like sysrescuecd) from the machine you want to image and give the command:

dd if=/dev/sda | gzip -1 - | ssh user@hostname dd of=image.gz

Assuming sda is your hard drive. This sends the local disks data to the remote machine. To restore the image boot knoppix on the machine to restore and pull the image that you created and dump it back with the command:

ssh user@hostname dd if=image.gz | gunzip -1 - | dd of=/dev/sda

php strict warnings | Drupal Groups

2013-06-19T12:10:00.001+02:00

php strict warnings | Drupal Groups:

Thanks, forcing 5.2 worked.

Disabling warnings from php strict in Drupal

Posted by DarrellDuane on September 26, 2012 at 3:21pm

I'm well aware that hacking core is frowned upon, but sometimes ya gotta do what ya gotta do. Running drupal 6 with views 2.x using PHP 5.4 (Fedora 17 in this case) yields a number of strict warnings that are not easily turned off. PHP doesn't allow you to turn off these strict warnings in its php.ini file, they come through even if E_STRICT is disabled, see
http://stackoverflow.com/questions/4692999/e-strict-messages-thrown-thou... for more details.
In order to fix this, I have edited includes/bootstrap.inc and put the following at the beginning of function drupal_set_message starting after line 991:

// filter out strict warnings due to conflicts between views-2.x and PHP 5.4
if( !strncmp($message, 'strict warning:', 14) ) {
return isset($_SESSION['messages']) ? $_SESSION['messages'] : NULL;
}

There may be other messages that come through that you want to filter in this manner as well.
If there is a hook that can be used to filter drupal messages or a cleaner way to do this by adding a module instead of hacking core I'd love to hear about it.

You and Your Research

2013-02-22T18:09:00.001+01:00

You and Your Research: "r. Somehow they seem to work on slightly the wrong thing - not much, but enough that they miss fame."

'via Blog this'

Ubuntu 11.04 and perlbrew installation problems · Issue #64 · gugod/App-perlbrew · GitHub

2012-09-21T09:07:00.001+02:00

Those who have a problem compiling old versions of Perl on Ubuntu, might want to read the next article.

Ubuntu 11.04 and perlbrew installation problems · Issue #64 · gugod/App-perlbrew · GitHub:

'via Blog this'

VM Farms: Cross Process Locking and Synchronization in Python

2012-09-11T09:28:00.001+02:00

Named locking in Python

VM Farms: Cross Process Locking and Synchronization in Python:

'via Blog this'

Julius • View topic - Using a SRILM N-gram on Julius

2012-07-30T15:20:00.001+02:00

This is very important for those using Julius ASR. This is a must.

Julius • View topic - Using a SRILM N-gram on Julius:

'via Blog this'

Introduction to ASR and HTK in particular

2012-07-25T10:00:00.001+02:00

Introduction to ASR and HTK in particular

'via Blog this'

USING HTK

2012-07-24T10:51:00.001+02:00

USING HTK:

UNDERSTANDING HTK ERROR MESSAGES

'via Blog this'

A nice tutorial for HTK

2012-07-19T15:39:00.001+02:00

A nice tutorial about the HTK : http://my.fit.edu/~vkepuska/HTK/icassp06_tutorial.pdf

'via Blog this'

Information Theory, Pattern Recognition, and Neural Networks

2012-07-19T15:06:00.001+02:00

One of the most interesting courses ever ;-)

Watch the lectures 9 - 14!

Information Theory, Pattern Recognition, and Neural Networks:

'via Blog this'

How do I sync multiple google calendars to my iPad.. It will only sync one. - Google Groups

2012-04-22T22:15:00.001+02:00

How do I sync multiple google calendars to my iPad.. It will only sync one. - Google Groups:

Use this: www.google.com/calendar/iphoneselect

'via Blog this'

Connection between MaxEnt and Logistic Regression

2012-01-28T21:54:00.001+01:00

This is worth reading.

McDonald, Ryan. 2007. Generalized Linear Classifiers in NLP. Course presented to the Swedish Graduate School in Language Technology.