Skip to content

Latest commit

 

History

History
275 lines (194 loc) · 14.8 KB

README.rst

File metadata and controls

275 lines (194 loc) · 14.8 KB

CorporaCreator

This is a command line tool to create Common Voice corpora.

After checking this repo out one installs the corresponding python package as follows

CorporaCreator$ python3 setup.py install

Given the clips.tsv file dumped from the Common Voice database, you can create a corpus (for each language in the clips.tsv file) as follows:

CorporaCreator$ create-corpora -d corpora -f clips.tsv

This will create the corpora in the directory corpora from the clips.tsv file.

If you would like to just create corpora for a some language(s), you can pass the --langs flag as follows:

CorporaCreator$ create-corpora -d corpora -f clips.tsv --langs en fr

This will create the corpora only for English and French.

Each created corpus will contain the files valid.tsv, containing the validated clips; invalid.tsv, containing the invalidated clips; and other.tsv, containing clips that don't have sufficient votes to be considered valid or invalid. In addition it will contain the files train.tsv, the valid clips in the training set; dev.tsv, the valid clips in the validation set; and test.tsv, the valid clips in test set.

The split of valid.tsv into train.tsv, dev.tsv, and test.tsv is done such that the number of clips in dev.tsv or test.tsv is a "statistically significant" sample relataive to the number of clips in train.tsv. More specificially, if the population size is the number of clips in train.tsv, then the number of clips in dev.tsv or test.tsv is the sample size required for a confidence level of 99% and a margin of error of 1% for the train.tsv population size.

By default no sentence occurs more than once in train.tsv, dev.tsv, and test.tsv. However, one can relax this constraint using the -s command line parameter. The value of -s is the number of repeats allows for a sentence. So, for example, if one wanted to allow for a sentence to occur 3 times in a corpus, then one could use

CorporaCreator$ create-corpora -d corpora -f clips.tsv -s 3

With or without the use of the -s command line parameter, the result of running create-corpora will be a directory containing the following files:

CorporaCreator$ tree corpora
corpora
├── br
│   ├── dev.tsv
│   ├── invalid.tsv
│   ├── other.tsv
│   ├── test.tsv
│   ├── train.tsv
│   └── valid.tsv
├── ca
│   ├── dev.tsv
│   ├── invalid.tsv
│   ├── other.tsv
│   ├── test.tsv
│   ├── train.tsv
│   └── valid.tsv
├── cnh
│   ├── dev.tsv
│   ├── invalid.tsv
│   ├── other.tsv
│   ├── test.tsv
│   ├── train.tsv
│   └── valid.tsv
.
.
.
├── tt
│   ├── dev.tsv
│   ├── invalid.tsv
│   ├── other.tsv
│   ├── test.tsv
│   ├── train.tsv
│   └── valid.tsv
└── zh-TW
    ├── dev.tsv
    ├── invalid.tsv
    ├── other.tsv
    ├── test.tsv
    ├── train.tsv
    └── valid.tsv

19 directories, 114 files

The purpose of the create-corpora command line tool is to provide a jumping-off point for contributors. The data in the alpha release of the Common Voice data is, unfortunately, in need of cleaning and the create-corpora command line tool provides a plugin for each language that allows for the language communities to aid in cleaning the data.

The clips.tsv file is a tab separated file containing a dump of the raw data from Common Voice with the following columns:

  1. client_id - A unique identifier for the contributor that was randomly generated when the contributor joined
  2. path - The path to the audio file containing the contribution
  3. sentence - The sentence the contributor was asked to read
  4. up_votes - The number of up votes for the contribution
  5. down_votes - The number of down votes for the contribution
  6. age - The age range of the contributor, if the contributor reported it
  7. gender - The gender of the contributor, if the contributor reported it
  8. accents - The accent of the contributor, if the contributor reported it
  9. variant - The variant of the language that contributor speaks, if the contributor reported it
  10. locale - The locale describing the language the contributor was reading
  11. segment - Shows whether the sentence belongs to a specific segment
  12. sentence_domain - The domain the sentence belongs to
  13. bucket - The "bucket" (train, dev, or test) the clip is currently assigned to

Our problem is that data in the column sentence needs to be cleaned, as there are various problems with the data in the sentence column. For example, some sentences contain HTML fragments. Some contain spelling errors. Some contain digits, e.g. "Room 4025" that allow for many valid readings. Some contain errors which we at Mozilla are not even aware of.

To actually see what needs to be cleaned first hand, the best thing to do is to run create-corpora as suggested above:

CorporaCreator$ create-corpora -d corpora -f clips.tsv

which will create the corpora in the directory corpora from the clips.tsv file. Then examine, for English say, the file corpora/en/valid.tsv to see which sentences there need cleaning. For other languages you would examine the corresponding file, e.g. for French it would be corpora/fr/valid.tsv.

To correct these problems we outfitted create-corpora with a plugin common.py that is responsible for cleaning sentences in a language independent manner. For example, if a sentence contains HTML fragments, then the HTML fragments would be removed by common.py.

The language independent cleaning is done by the common() method in common.py:

def common(sentence):
    """Cleans up the passed sentence in a language independent manner, removing or reformatting invalid data.
    Args:
      sentence (str): Sentence to be cleaned up.
    Returns:
      (boolean,str): A boolean indicating validity and cleaned up sentence.
    """
    ...
    # Clean sentence in a language independent manner
    ...
    return is_valid, sentence

This method is input the sentence to clean, cleans the sentence in a language independent manner, and returns the cleaned sentence along with a boolean indicating its validity.

If the sentence is not able to be cleaned, e.g. it consisted only of HTML fragments, this method can return is_valid set to False.

Currently common.py decodes any URL encoded elements of sentence, removes any HTML tags in a sentence, removes any non-printable characters in a sentence, and marks as invalid any sentence containing digits, in that order. (For the details refer to common.py .) This seems to catch most language independent problems, but if you see more, please open an issue or make a pull request.

In addition to the language independent plugin common.py create-corpora can also support language-dependent cleaning. In order to add language-dependent cleaning, create a plugin named LOCALE.py in the preprocessors folder with a function definition also named LOCALE, where LOCALE is whatever ISO language-code is. NOTE: hyphens are not supported, so something like zh-TW would be named zhTW.py.

For example, the cleaning for English would be done by the en() method in a file named en.py:

def en(client_id, sentence):
    """Cleans up the passed sentence, removing or reformatting invalid data.
    Args:
      client_id (str): Client ID of sentence's speaker
      sentence (str): Sentence to be cleaned up.
    Returns:
      (str): Cleaned up sentence. Returning None or a `str` of whitespace flags the sentence as invalid.
    """
    # TODO: Clean up en data
    return sentence

This method accepts the sentence to clean along with the client_id of the contributor who read the sentence. It then cleans the sentence in a language dependent manner and returns the cleaned sentence. For a more complex example of what this could look like, refer to preprocessors/de.py.

If the sentence is not able to be cleaned, e.g. it is so mangled that it is impossible to determine how to correct it to a valid English sentence, this method can return None or a string containing only whitespace to indicate the sentence was invalid to begin with.

Of note is that in the language dependent case the method that does the cleaning takes not only the sentence but also the client_id of the contributor who read the sentence. In the language independent case this client_id was not present. However, for the language dependent case it's unfortunately required.

A sentence may contain text which is able to be read in many different, but valid, ways. For example, the sentence "I am in room 4025." can be validly read as "I am in room four oh two five". Equivalently, a valid reading is: "I am in room four zero two five". There are also other valid readings: "I am in room forty twenty five.", "I am in room four thousand twenty five."... To actually determine which of these readings a particular contributor gave, you have to listen to the audio, determine what they said, then replace the digits with text reflecting the contributor's reading, returning this cleaned sentence.

If you are interested in helping clean sentences for a particular language, or even cleaning in a language independent manner in common.py you can make a pull request that includes your changes. Here we will look at some common ways to correct sentences.

Suppose you found that one, or more English sentences had a misspelling of the word "masquerade" as "masqurade" (sic). As this is concerned with the English language you would write code in the en.py plugin. A simple solution would be to replace all occurrences of "masqurade" (sic) with "masquerade" in every sentence. One could do this as follows:

def en(client_id, sentence):
    """Cleans up the passed sentence, removing or reformatting invalid data.
    Args:
      client_id (str): Client ID of sentence's speaker
      sentence (str): Sentence to be cleaned up.
    Returns:
      (str): Cleaned up sentence. Returning None or a `str` of whitespace flags the sentence as invalid.
    """
    sentence = sentence.replace("masqurade", "masquerade")
    # TODO: Clean up en data
    return sentence

what you have to be careful about, and which is a complexity that this simple example ignores, is that the word you are replacing can not appear in a context where the replacement is invalid. For example, if "the" were mistyped as "teh", then doing the same replacement of "teh" with "the" would run the risk of converting "tehran" to "theran", an undesired consequence. So you have to be careful.

Suppose you found that one, or more English sentences used the abbreviation "STT" for "speech-to-text". Some people may have read "STT" as the letters "S T T". However, some may have known the abbreviation and read this as "speech-to-text". To determine which was done you have to hear the audio for each reading and write code that handles each contributor individually.

One could do this as follows:

def en(client_id, sentence):
    """Cleans up the passed sentence, removing or reformatting invalid data.
    Args:
      client_id (str): Client ID of sentence's speaker
      sentence (str): Sentence to be cleaned up.
    Returns:
      (str): Cleaned up sentence. Returning None or a `str` of whitespace flags the sentence as invalid.
    """
    if client_id == "8d59b8879856":
        sentence = sentence.replace("STT", "speech-to-text")
    if client_id == "48f3620be0fa":
        sentence = sentence.replace("STT", "S T T")
    # TODO: Clean up en data
    return sentence

To actually hear the audio, you have to request the audio from Mozilla. (See the information distributed with the alpha release as to how to obtain the audio.)

Once you have obtained the audio, you can hear the audio for a given sentence and client_id pair by finding the row corresponding to the sentence + client_id pair in clips.tsv, finding the path in that row, then playing the file corresponding to the row's path in the downloaded audio.

Suppose you found that one, or more English sentences used the text "room 4025". Some people may have read "room 4025" as "room four oh two five", some as "room four zero two five", some in a completely different way. Again, to determine which way the digits were read, you have to hear the audio for each reading and write code that handles each contributor individually.

One could do this as follows:

def en(client_id, sentence):
    """Cleans up the passed sentence, removing or reformatting invalid data.
    Args:
      client_id (str): Client ID of sentence's speaker
      sentence (str): Sentence to be cleaned up.
    Returns:
      (str): Cleaned up sentence. Returning None or a `str` of whitespace flags the sentence as invalid.
    """
    if client_id == "8d59b8879856":
        sentence = sentence.replace("room 4025", "room four oh two five")
    if client_id == "48f3620be0fa":
        sentence = sentence.replace("room 4025", "room four zero two five")
    # TODO: Clean up en data
    return sentence

To actually hear the audio, you have to request the audio from Mozilla. (See the information distributed with the alpha release as to how to obtain the audio.)

As in the case of abbreviations, you can hear the audio for a given sentence and client_id pair by finding the row corresponding to the sentence + client_id pair in clips.tsv, finding the path in that row, then playing the file corresponding to the row's path in the downloaded audio.