Adapting the default acoustic model



This page describes how to do some simple acoustic model adaptation to improve speech recognition in your configuration. Please note that the adaptation doesn’t necessary adapt for a particular speaker. It just improves the fit between the adaptation data and the model. For example you can adapt to your own voice to make dictation good, but you also can adapt to your particular recording environment, your audio transmission channel, your accent or accent of your users. You can use a model trained with clean broadcast data and telephone data to produce a telephone acoustic model by doing adaptation. Cross-language adaptation also make sense, you can for example adapt an English model to sounds of another language by creating a phoneset map and creating another language dictionary with an English phoneset.

The adaptation process takes transcribed data and improves the model you already have. It’s more robust than training and could lead to good results even if your adaptation data is small. For example, it’s enough to have 5 minutes of speech to significantly improve the dictation accuracy by adapting to the particular speaker.

The methods of adaptation are a bit different between PocketSphinx and Sphinx4 due to the different types of acoustic models used. For more technical information on that read the article about Acoustic Model Types.

Building the tools

You’ll minimally need PocketSphinx and SphinxTrain here. Build and install PocketSphinx following the instructions in the tutorial. SphinxTrain can be built using the same instructions:

cmake -S . -B build -DCMAKE_INSTALL_PREFIX=$HOME/cmusphinx -G Ninja
cmake --build build --target install

The tools will be located in $HOME/cmusphinx/libexec/sphinxtrain.

Creating an adaptation corpus

Now, you will need to create a corpus of adaptation data. The corpus will consist of

  • a list of sentences
  • a dictionary describing the pronunciation of all the words in that list of sentences
  • a recording of you speaking each of those sentences

Required files

The actual set of sentences you use is somewhat arbitrary, but ideally it should have good coverage of the most frequently used words or phonemes in the set of sentences or the type of text you want to recognize. For example, if you want to recognize isolated commands, you need tor record them. If you want to recognize dictation, you need to record full sentences. For simple voice adaptation we have had good results simply by using sentences from the CMU ARCTIC text-to-speech databases. To that effect, here are the first 20 sentences from ARCTIC, a .fileids file, and a transcription file:

The sections below will refer to these files, so, if you want to follow along we recommend downloading these files now. You should also make sure that you have downloaded and compiled sphinxbase and sphinxtrain.

Recording your adaptation data

In case you are adapting to a single speaker you can record the adaptation data yourself. This is unfortunately a bit more complicated than it ought to be.
Basically, you need to record a single audio file for each sentence in the adaptation corpus, naming the files according to the names listed in arctic20.transcription and arctic20.fileids.

In addition, you need to make sure that you record at a sampling rate of 16 kHz (or 8 kHz if you adapt a telephone model) in mono with a single channel.

The simplest way would be to start a sound recorder like Audacity or Wavesurfer and read all sentences in one big audio file. Then you can cut the audio files on sentences in a text editor and make sure every sentence is saved in the corresponding file. The file structure should look like this:

arctic_0001.wav
arctic_0002.wav
.....
arctic_0019.wav
arctic20.fileids
arctic20.transcription

You should verify that these recordings sound okay. To do this, you can play them back with:

for i in *.wav; do play $i; done

If you already have a recording of the speaker, you can split it on sentences and create the .fileids and the .transcription files.

If you are adapting to a channel, accent or some other generic property of the audio, then you need to collect a little bit more recordings manually. For example, in a call center you can record and transcribe hundred calls and use them to improve the recognizer accuracy by means of adaptation.

Adapting the acoustic model

First we will copy the default acoustic model from PocketSphinx into the current directory in order to work on it. Assuming that you installed PocketSphinx under $HOME/cmusphinx, the acoustic model directory is $HOME/cmusphinx/share/pocketsphinx/model/en-us/en-us. Copy this directory to your working directory:

cp -a $HOME/cmusphinx/share/pocketsphinx/model/en-us/en-us .

Let’s also copy the dictionary and the langauge model for testing:

cp -a $HOME/cmusphinx/share/pocketsphinx/model/en-us/cmudict-en-us.dict .
cp -a $HOME/cmusphinx/share/pocketsphinx/model/en-us/en-us.lm.bin .

Generating acoustic feature files

In order to run the adaptation tools, you must generate a set of acoustic model feature files from these WAV audio recordings. This can be done with the sphinx_fe tool from SphinxBase. It is imperative that you make sure you are using the same acoustic parameters to extract these features as were used to train the standard acoustic model. Since PocketSphinx 0.4, these are stored in a file called feat.params in the acoustic model directory. You can simply add it to the command line for sphinx_fe, like this:

$HOME/cmusphinx/libexec/sphinxtrain/sphinx_fe -argfile en-us/feat.params \
        -samprate 16000 -c arctic20.fileids \
       -di . -do . -ei wav -eo mfc -mswav yes

You should now have the following files in your working directory:

en-us
arctic_0001.mfc
arctic_0001.wav
arctic_0002.mfc
arctic_0002.wav
arctic_0003.mfc
arctic_0003.wav
.....
arctic_0020.wav
arctic20.fileids
arctic20.transcription
cmudict-en-us.dict
en-us.lm.bin

Converting the sendump and mdef files

Some models like en-us are distributed in compressed version. Extra files that are required for adaptation are excluded to save space. For the en-us model from pocketsphinx you can download the full version suitable for adaptation:

cmusphinx-en-us-ptm-5.2.tar.gz

Make sure you are using the full model with the mixture_weights file present.

If the mdef file inside the model is converted to binary, you will also need to convert the mdef file from the acoustic model to the plain text format used by the SphinxTrain tools. To do this, use the pocketsphinx_mdef_convert program:

$HOME/cmusphinx/bin/pocketsphinx_mdef_convert -text en-us/mdef en-us/mdef.txt

In the downloads the mdef is already in the text form.

Accumulating observation counts

The next step in the adaptation is to collect statistics from the adaptation data. This is done using the bw program from SphinxTrain. You should be able to find the bw tool in a sphinxtrain installation in the folder $HOME/cmusphinx/libexec/sphinxtrain.

Now, to collect the statistics, run:

$HOME/cmusphinx/libexec/sphinxtrain/bw \
 -hmmdir en-us \
 -moddeffn en-us/mdef.txt \
 -ts2cbfn .ptm. \
 -feat 1s_c_d_dd \
 -svspec 0-12/13-25/26-38 \
 -cmn current \
 -agc none \
 -dictfn cmudict-en-us.dict \
 -ctlfn arctic20.fileids \
 -lsnfn arctic20.transcription \
 -accumdir .

Make sure the arguments in the bw command match the parameters in the feat.params file inside the acoustic model folder. Please note that not all the parameters from feat.param are supported by bw. bw for example doesn’t suppport upperf or other feature extraction parameters. You only need to use parameters which are accepted, other parameters from feat.params should be skipped.

For example, for a continuous model you don’t need to include the svspec option. Instead, you need to use just -ts2cbfn .cont. For semi-continuous models use -ts2cbfn .semi. If the model has a feature_transform file like the en-us continuous model, you need to add the -lda feature_transform argument to bw, otherwise it will not work properly.

If you are missing the noisedict file, you also need an extra step. Copy the fillerdict file into the directory that you choose in the hmmdir parameter and renaming it to noisedict.

Creating a transformation with MLLR

MLLR transforms are supported by pocketsphinx and sphinx4. MLLR is a cheap adaptation method that is suitable when the amount of data is limited. It’s a good idea to use MLLR for online adaptation. MLLR works best for a continuous model. Its effect for semi-continuous models is very limited since semi-continuous models mostly rely on mixture weights. If you want the best accuracy you can combine MLLR adaptation with MAP adaptation below. On the other hand, because MAP requires a lot of adaptation data it is not really practical to use it for continuous models. For continuous models MLLR is more reasonable.

Next, we will generate an MLLR transformation which we will pass to the decoder to adapt the acoustic model at run-time. This is done with the mllr_solve program:

$HOME/cmusphinx/libexec/sphinxtrain/mllr_solve \
    -meanfn en-us/means \
    -varfn en-us/variances \
    -outmllrfn mllr_matrix -accumdir .

This command will create an adaptation data file called mllr_matrix. Now, if you wish to decode with the adapted model, simply add -mllr mllr_matrix (or whatever the path to the mllr_matrix file you created is) to your pocketsphinx command line.

Updating the acoustic model files with MAP

MAP is a different adaptation method. In this case, unlike for MLLR, we don’t create a generic transform but update each parameter in the model. We will now copy the acoustic model directory and overwrite the newly created directory with the adapted model files:

cp -a en-us en-us-adapt

To apply the adaptation, use the map_adapt program:

$HOME/cmusphinx/libexec/sphinxtrain/map_adapt \
    -moddeffn en-us/mdef.txt \
    -ts2cbfn .ptm. \
    -meanfn en-us/means \
    -varfn en-us/variances \
    -mixwfn en-us/mixture_weights \
    -tmatfn en-us/transition_matrices \
    -accumdir . \
    -mapmeanfn en-us-adapt/means \
    -mapvarfn en-us-adapt/variances \
    -mapmixwfn en-us-adapt/mixture_weights \
    -maptmatfn en-us-adapt/transition_matrices

Recreating the adapted sendump file

If you want to save space for the model you can use a sendump file which is supported by PocketSphinx. For Sphinx4 you don’t need that. To recreate the sendump file from the updated mixture_weights file run:

$HOME/cmusphinx/libexec/sphinxtrain/mk_s2sendump \
    -pocketsphinx yes \
    -moddeffn en-us-adapt/mdef.txt \
    -mixwfn en-us-adapt/mixture_weights \
    -sendumpfn en-us-adapt/sendump

Congratulations! You now have an adapted acoustic model.

The en-us-adapt/mixture_weights and en-us-adapt/mdef.txt files are not used by the decoder, so, if you like, you can delete them to save some space.

Other acoustic models

For Sphinx4, the adaptation is the same as for PocketSphinx, except that Sphinx4 can not read the binary compressed mdef and sendump files, you need to leave the mdef and the mixture weights file.

Testing the adaptation

After you have done the adaptation, it’s critical to test the adaptation quality. To do that you need to setup the database similar to the one used for adaptation. To test the adaptation you need to configure the decoding with the required paramters, in particular, you need to have a language model <your.lm>. For more details see the tutorial on Building a Language Model. The detailed process of testing the model is covered in another part of the tutorial.

You can try to run the decoder on the original acoustic model and on the new acoustic model to estimate the improvement.

Using the model

After adaptation, the acoustic model is located in the folder en-us-adapt. You need only that folder. The model should have the following files:

mdef
feat.params
mixture_weights
means
noisedict
transition_matrices
variances

depending on the type of the model you trained.

To use the model in PocketSphinx, simply put the model files to the resources of your application. Then point to it with the -hmm option:

pocketsphinx -hmm `<your_new_model_folder>` -lm `<your_lm>` \
    -dict `<your_dict>` single test.wav

To use the trained model in Sphinx4, you need to update the model location in the code.

Troubleshooting

If the adaptation didn’t improve your results, first test the accuracy and make sure it’s good.

I have no idea where to start looking for the problem…

  1. Test whether the accuracy on the adaptation set improved
  2. Accuracy improved on adaptation set ⇢ check if your adaptation set matches with your test set
  3. Accuracy didn’t improve on adaptation set ⇢ you made a mistake during the adaptation

…or how much improvement I might expect through adaptation

From few sentences you should get about 10% relative WER improvement.

I’m lost about…

…whether it needs more/better training data, whether I’m not doing the adaptation correctly, whether my language model is the problem here, or whether there is something intrinsically wrong with my configuration.

Most likely you just ignored some error messages that were printed to you. You obviosly need to provide more information and give access to your experiment files in order to get more definite advise.

What’s next

We hope the adapted model gives you acceptable results. If not, try to improve your adaptation process by:

  1. Adding more adaptation data
  2. Adapting your language mode / using a better language model

Building a language model Training an acoustic model