Building a language model

The language model is an important component of the configuration which tells the decoder which sequences of words are possible to recognize.

There are several types of models: keyword lists, grammars and statistical language models and phonetic language models. They have different capabilities and performance properties. You can chose any decoding mode according to your needs and you can even switch between modes in runtime. See the Pocketsphinx tutorial for more details.

Keyword lists

Pocketsphinx supports a keyword spotting mode where you can specify a list of keywords to look for. The advantage of this mode is that you can specify a threshold for each keyword so that keywords can be detected in continuous speech. All other modes will try to detect the words from a grammar even if you used words which are not in the grammar. A typical keyword list looks like this:

oh mighty computer /1e-40/
hello world /1e-30/
other phrase /1e-20/

The threshold must be specified for every keyphrase. For shorter keyphrases you can use smaller thresholds like 1e-1, for longer keyphrases the threshold must be bigger, up to 1e-50. If your keyphrase is very long – larger than 10 syllables – it is recommended to split it and spot for parts separately. The threshold must be tuned to balance between false alarms and missed detections. The best way to do this is to use a prerecorded audio file. The common tuning process is the following:

  1. Take a long recording with few occurrences of your keywords and some other sounds. You can take a movie sound or something else. The length of the audio should be approximately 1 hour.
  2. Run a keyword spotting on that file with different thresholds for every keyword, use the following command:
     pocketsphinx_continuous -infile <your_file.wav> -keyphrase <your keyphrase> \
      -kws_threshold <your_threshold> -time yes

    The command will print many lines, some of them are keywords with detection times and confidences. You can also disable extra logs with the -logfn your_file.log option to avoid clutter.

  3. From your keyword spotting results count how many false alarms and missed detections you’ve encountered.
  4. Select the threshold with the smallest amount of false alarms and missed detections.

For the best accuracy it is better to have a keyphrase with 3-4 syllables. Too short phrases are easily confused.

Keyword lists are only supported by pocketsphinx, sphinx4 cannot handle them.

Using keyword lists with PocketSphinx

To use keyword list in the command line specify it with the -kws option. You can also use a -keyphrase option to specify a single keyphrase.

In Python you can either specify options in the configuration object or add a named search for a keyphrase:

decoder.set_kws('keyphrase', kws_file)

In Android it looks similar:

recognizer.setKws('keyphrase', kwsFile);

Please note that -kws conflicts with the -lm and -jsgf options. You cannot specify both.


A grammar describes a very simple type of the language for command and control. They are usually written by hand or generated automatically within the code. Grammars usually do not have probabilities for word sequences, but some elements might be weighed. They can be created with the Java Speech Grammar Format (JSGF) and usually have a file extension like .gram or .jsgf.

Grammars allow you to specify possible inputs very precisely, for example, that a certain word might be repeated only two or three times. However, this strictness might be harmful if your user accidentally skips the words which the grammar requires. In that case the whole recognition will fail. For that reason it is better to make grammars more flexible. Instead of phrases, just list the bag of words allowing arbitrary order. Avoid very complex grammars with many rules and cases. It just slows down the recognizer and you can use simple rules instead. In the past, grammars required a lot of effort to tune them, to assign variants properly and so on. The big VXML consulting industry was about that.

Building a grammar

Grammars are usually written manually in the Java Speech Grammar Format (JSGF):

#JSGF V1.0;

grammar hello;
public <greet> = (good morning | hello) ( bhiksha | evandro | rita | will );

For more information on JSGF see the full documentation on W3C.

Using your grammar with PocketSphinx

To use your grammar in the command line specify it with the -jsgf option.

In Python you can either specify options in the configuration object or add a named search for a grammar:

decoder.set_jsgf('grammar', jsgf_file)

In Android this looks similiar:

recognizer.setJsgf('grammar', jsgfFile);

Please note that -jsgf conflicts with the -kws and -jsgf options. You cannot specify both.

Language models

Statistical language models describe more complex language. They contain probabilities of the words and word combinations. Those probabilities are estimated from sample data and automatically have some flexibility. Every combination from the vocabulary is possible, although the probability of each combination will vary. For example, if you create a statistical language model from a list of words it will still allow to decode word combinations even though this might not have been your intent.

Overall, statistical language models are recommended for free-form input where the user could say anything in a natural language. They require way less engineering effort than grammars. You just list the possible sentences. For example, you might list numbers like “twenty one” and “thirty three” and a statistical language model will allow “thirty one” with a certain probability as well.

In general, modern speech recognition interfaces tend to be more natural and avoid the command-and-control style of the previous generation. For that reason most interface designers prefer natural language recognition with a statistical language model instead of using old-fashioned VXML grammars.

On the topic of desiging VUI interfaces you might be interested in the following book: It’s Better to Be a Good Machine Than a Bad Person: Speech Recognition and Other Exotic User Interfaces at the Twilight of the Jetsonian Age by Bruce Balentine.

There are many ways to build statistical language models. When your data set is large, it makes sense to use the CMU language modeling toolkit. When a model is small, you can use a quick online web service. When you need specific options or you just want to use your favorite toolkit which builds ARPA models, you can use this as well.

A language model can be stored and loaded in three different formats: text ARPA format, binary BIN format and binary DMP format. The ARPA format takes more space but it is possible to edit it. ARPA files have an .lm extension. Binary formats take significantly less space and load faster. Binary files have a .lm.bin extension. It is also possible to convert between these formats. The DMP format is obsolete and not recommended.

Building a statistical language model

Text preparation

First of all you need to prepare a large collection of clean texts. Expand abbreviations, convert numbers to words, clean non-word items. For example to clean Wikipedia XML dumps you can use special Python scripts like Wikiextractor. To clean HTML pages you can try BoilerPipe. It’s a nice package specifically created to extract text from HTML.

For an example on how to create a language model from Wikipedia text, please see this blog post. Movie subtitles are also a good source for spoken language.

Once you have gone through the language modeling process, please submit your language model to the CMUSphinx project. We’ll be happy to share it!

Language modeling for Mandarin and other similar languages, is largely the same as for English, with one additional consideration. The difference is that the input text must be word segmented. A segmentation tool and an associated word list is provided to accomplish this.

Training an ARPA model with SRILM

Training a model with the SRI Language Modeling Toolkit (SRILM) is easy. That’s why we recommend it. Moreover, SRILM is the most advanced toolkit up to date. To train a model you can use the following command:

ngram-count -kndiscount -interpolate -text train-text.txt -lm your.lm

You can prune the model afterwards to reduce the size of the model:

ngram -lm your.lm -prune 1e-8 -write-lm your-pruned.lm

After training it is worth it to test the perplexity of the model on the test data:

ngram -lm your.lm -ppl test-text.txt

Training an ARPA model with CMUCLMTK

You need to download and install the language model toolkit for CMUSphinx (CMUCLMTK). See the download page for details.

The process for creating a language model is as follows:

1) Prepare a reference text that will be used to generate the language model. The language model toolkit expects its input to be in the form of normalized text files, with utterances delimited by <s> and </s> tags. A number of input filters are available for specific corpora such as Switchboard, ISL and NIST meetings, and HUB5 transcripts. The result should be the set of sentences that are bounded by the start and end markers of the sentence: <s> and </s>. Here’s an example:

<s> generally cloudy today with scattered outbreaks of rain and drizzle
persistent and heavy at times </s>
<s> some dry intervals also with hazy sunshine especially in eastern parts in
the morning </s>
<s> highest temperatures nine to thirteen Celsius in a light or moderate mainly
east south east breeze </s>
<s> cloudy damp and misty today with spells of rain and drizzle in most places
much of this rain will be light and patchy but heavier rain may develop in the
west later </s>

More data will generate better language models. The weather.txt file from sphinx4 (used to generate the weather language model) contains nearly 100,000 sentences.

2) Generate the vocabulary file. This is a list of all the words in the file:

 text2wfreq < weather.txt | wfreq2vocab > weather.tmp.vocab

3) You may want to edit the vocabulary file to remove words (numbers, misspellings, names). If you find misspellings, it is a good idea to fix them in the input transcript.

4) If you want a closed vocabulary language model (a language model that has no provisions for unknown words), then you should remove sentences from your input transcript that contain words that are not in your vocabulary file.

5) Generate the ARPA format language model with the commands:

  text2idngram -vocab weather.vocab -idngram weather.idngram < weather.closed.txt
  idngram2lm -vocab_type 0 -idngram weather.idngram -vocab weather.vocab -arpa weather.lm

6) Generate the CMU binary form (BIN):

 sphinx_lm_convert -i weather.lm -o weather.lm.bin

Building a simple language model using a web service

If your language is English and the text is small it’s sometimes more convenient to use a web service to build it. Language models built in this way are quite functional for simple command and control tasks. First of all you need to create a corpus.

The “corpus” is just a list of sentences that you will use to train the language model. As an example, we will use a hypothetical voice control task for a mobile Internet device. We’d like to tell it things like “open browser”, “new e-mail”, “forward”, “backward”, “next window”, “last window”, “open music player”, and so forth. So, we’ll start by creating a file called corpus.txt:

open browser
new e-mail
next window
last window
open music player

Then go to the LMTool page. Simply click on the “Browse…” button, select the corpus.txt file you created, then click “COMPILE KNOWLEDGE BASE”.

You should see a page with some status messages, followed by a page entitled “Sphinx knowledge base”. This page will contain links entitled “Dictionary” and “Language Model”. Download these files and make a note of their names (they should consist of a 4-digit number followed by the extensions .dic and .lm). You can now test your newly created language model with PocketSphinx.

Using other language model toolkits

There are many toolkits that create an ARPA n-gram language model from text files.

Some toolkits you can try:

If you are training a large vocabulary speech recognition system, the language model training is outlined in a separate page about large scale language models.

Once you have created an ARPA file you can convert the model to a binary format for faster loading.

Converting a model into the binary format

To quickly load large models you probably would like to convert them to a binary format that will save your decoder initialization time. That’s not necessary with small models. Pocketsphinx and sphinx3 can handle both of them with the -lm option. Sphinx4 automatically detects the format by the extension of the lm file.

The ARPA format and BINARY format are mutually convertable. You can produce the other file with the sphinx_lm_convert command from sphinxbase:

sphinx_lm_convert -i model.lm -o model.lm.bin
sphinx_lm_convert -i model.lm.bin -ifmt bin -o model.lm -ofmt arpa

You can also convert old DMP models to a binary format this way.

In the next section we will deal with how to use, test, and improve the language model you created.

Using your language model with PocketSphinx

If you have installed PocketSphinx, you will have a program called pocketsphinx_continuous which can be run from the command line to recognize speech. Assuming it is installed under /usr/local, and your language model and dictionary are called 8521.dic and 8521.lm and placed in the current folder, try running the following command:

pocketsphinx_continuous -inmic yes -lm 8521.lm -dict 8521.dic

This will use your new language model, the dictionary and the default acoustic model. On Windows you also have to specify the acoustic model folder with the -hmm option:

bin/Release/pocketsphinx_continuous.exe -inmic yes -lm 8521.lm -dict 8521.dic -hmm model/en-us/en-us

You will see a lot of diagnostic messages, followed by a pause, then the output “READY…“. Now you can try speaking some of the commands. It should be able to recognize them with full accuracy. If not, you may have problems with your microphone or sound card.

Using your language model with Sphinx4

In the Sphinx4 high-level API you need to specify the location of the language model in your Configuration:


If the model is in the resources you can reference it with "resource:URL":


Also see the Sphinx4 tutorial for more details.

Building a dictionary Adapting an existing acoustic model