Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time



A bidirectional recurrent neural network model with attention mechanism for restoring missing inter-word punctuation in unsegmented text.

The model can be trained in two stages (second stage is optional):

  1. First stage is trained on punctuation annotated text. Here the model learns to restore puncutation based on textual features only.
  2. Optional second stage can be trained on punctuation and pause annotated text. In this stage the model learns to combine pause durations with textual features and adapts to the target domain. If pauses are omitted then only adaptation is performed. Second stage with pause durations can be used for example for restoring punctuation in automatic speech recognition system output.

How well does it work?

Remember that all the scores given below are on unsegmented text and we did not use prosodic features, so, among other things, the model has to detect sentence boundaries in addition to the boundary type (?QUESTIONMARK, .PERIOD or !EXCLAMATIONMARK) based entirely on textual features. The scores are computed on the test set.

Training speed with default settings, an optimal Theano installation and a modern GPU should be around 10000 words per second.

Pretrained models can be downloaded here (Demo + 2 models from the Interspeech paper).

English TED talks

Training set size: 2.1M words. First stage only. More details can be found in this paper. For comparison, our previous model got an overall F1-score of 50.8.

,COMMA 64.4 45.2 53.1
?QUESTIONMARK 67.5 58.7 62.8
.PERIOD 72.3 71.5 71.9
Overall 68.9 58.1 63.1

English Europarl v7

Training set size: 40M words. First stage only. Details in ./example.

You can try to compete with this model here.

?QUESTIONMARK 77.7 73.2 75.4
,COMMA 68.9 72.0 70.4
-DASH 55.9 8.8 15.2
:COLON 60.9 23.8 34.2
;SEMICOLON 44.7 1.1 2.2
.PERIOD 84.7 84.1 84.4
Overall 75.7 73.9 74.8


  • Python 2.7
  • Numpy
  • Theano

Requirements for data:

  • Cleaned text files for training and validation of the first phase model. Each punctuation symbol token must be surrounded by spaces.

    Example: to be ,COMMA or not to be ,COMMA that is the question .PERIOD

  • (Optional) Pause annotated text files for training and validation of the second phase model. These should be cleaned in the same way as the first phase data. Pause durations in seconds should be marked after each word with a special tag <sil=0.200>. Punctuation mark, if any, must come after the pause tag.

    Example: to <sil=0.000> be <sil=0.100> ,COMMA or <sil=0.000> not <sil=0.000> to <sil=0.000> be <sil=0.150> ,COMMA that <sil=0.000> is <sil=0.000> the <sil=0.000> question <sil=1.000> .PERIOD

    Second phase data can also be without pause annotations to do just target domain adaptation.

Make sure that first words of sentences don't have capitalized first letters. This would give the model unfair hints about period locations. Also, the text files you use for training and validation must be large enough (at least minibatch_size x sequence_length of words, which is 128x50=6400 words with default settings), otherwise you might get an error.


Vocabulary size, punctuation tokens and their mappings, and converted data location can be configured in the header of Some model hyperparameters can be configured in the headings of and Learning rate and hidden layer size can be passed as arguments.


First step is data conversion. Assuming that preprocessed and cleaned *.train.txt, *.dev.txt and *.test.txt files are located in <data_dir>, the conversion can be initiated with:

python <data_dir>

If you have second stage data as well, then:

python <data_dir> <second_stage_data_dir>

The first stage can be trained with:

python <model_name> <hidden_layer_size> <learning_rate>

e.g python <model_name> 256 0.02 works well.

Second stage can be trained with:

python <model_name> <hidden_layer_size> <learning_rate> <first_stage_model_path>

Preprocessed text can be punctuated with e.g:

cat | python <model_path> <model_output_path>

or, if pause annotations are present in and you have a second stage model trained on pause annotated data, then:

cat | python <model_path> <model_output_path> 1

Punctuation tokens in don't have to be removed - the script ignores them.

Error statistics in this example can be computed with:

python <model_output_path>

You can play with a trained model with (assumes the input text is similarly preprocessed as the training data):

python <model_path>

or with:

python <model_path> 1

if you want to see, which words the model sees as UNKs (OOVs).


The software is described in:

  author    = {Ottokar Tilk and Tanel Alum{\"a}e},
  title     = {Bidirectional Recurrent Neural Network with Attention Mechanism for Punctuation Restoration},
  booktitle = {Interspeech 2016},
  year      = {2016}

We used the release v1.0 in the paper.