Corpus development

This chapter discusses the techniques used to design a good corpus for recording for use in general speech synthesis. The basic requirements for a speech synthesis corpus are:

To make life easier in designing prompt list we have included scripts which go some way to aid the process. The general idea is to take a very large amount of text and automatically find "nice" utterances that match these criteria.

The CMU ARCTIC Database prompt list [kominek 2003] was created very much in this way, though with an earlier version of the scripts.

As with most of our work there is a single script, that does a number of basic stages. The default is reasonable in many cases, but with prompt selection it is always worth hand checking and potentially modifying and refining the process.

The basic idea is to limit the chosen utterances to those of a reasonable length (5-15 words), only choose sentences with high frequency words (which should be easier to say and less ambiguous in pronunciation, also restrict to words that are in the lexicon (avoiding letter to sound issues).

The script make_nice_prompts is set up for two classes of language, Latin script languages and non-Latin script languages. Though as the non-Latin case is much more varied you may need to modify things. We have successfully used it for UTF-8 encoded Hindi.

For the Latin-script languages (as opposed "asis" cases) we downcase the text when looking for variation. Although some Latin based language make a significant case distinction, e.g. German, this is a reasonable route to avoid sentences with too many repeated words in them.

First gather lots of text data. When we say lots we mean at least millions of words, or even 10s on millions of words. This basic selection process is aimed at getting sentences for general voices and hence as large amount of starting data as possible is important.

Please also note the copyright of the data you are selecting from. In CMU ARCTIC we used out-of-copyright texts from the Gutenberg project, so there would be no issue in distributing the data. Copyright law in many countries allows for small subsets for copyright data, but this fair use is often argued by some. There may not actually be a good solution to this, News stories, are typically copyright by the press agency releasing them. Licenses on LDC data are often sufficient for using such texts to build prompts and then having no restriction on the voices generated, but the database itself may be under question. If you care about distribution (free or selling) you will need to address these issues.

The first stage once you have collected you data is check its encoding. Make sure its all the same encoding. Also check its reasonable. For example the Europal data is nice and clean (as conditioned for Machine Translation models) but the punctuation has been separated from the words. You may want to to de-htmlify your data before passing it to the selection routines.

Once you have a set of nice relatively clean data, you use the first stage of the script. This finds the word frequencies of all the tokens in the text data.

    $FESTVOXDIR/src/promptselect/make_nice_prompts find_freq TEXT0 TEXT1 TEXT2

Give all the text files as arguments to this script. The working files will be created in the current directory, but the text file arguments may be pathnames.

The next stage is to build a Festival lexicon for the most frequent words. By default we select the top 5000 words, which has proven a reasonable choice. You can optionally override the 5000 with an argument.

    $FESTVOXDIR/src/promptselect/make_nice_prompts make_freq_lex

The next stage processes each sentence in the large text database to find those utterances that are "nice". That of reasonable length, has only words in the frequency lexicon, no strange punctuation, capitals at the beginning, and punctuation at the end, and a few other heuristic rule conditions. These seem to work well for the Latin-script languages (though it is possible the conditions are overly strict for some languages).

Although Festival's text front end is used for processing the text, you do not need to build any language specific text front end (at least not normally). Finding nice prompts is considered one of the very first parts of buildings support for a new language, so we are aware that there will be almost no resources available for the target language yet.

Because this process is using Festival's front end, it is not fast, as it needs to process the whole text database. It is not unusual for this to take a number of hours to process. While processing "nice" utterances are written to You should check this regularly in case there is some inappropriate condition in the rules and you are getting the wrong type of data.

    $FESTVOXDIR/src/promptselect/make_nice_prompts find_nice TEXT0 TEXT1 ...

Note this will only search for the first 100,000 nice utterances, from the data set, you can change the number in the script if you want more (or less).

Once the "nice" utterances are found you can now find those nice utterances that have the best phonetic coverage. There are two mechanism available here. Because this is often the very first stage in building support for a new language, no lexicon and phonetics are available, thus selecting based on phonetic is not an option. Therefore we provide a simpler technique that selections based on letter coverage (in fact di-letter coverage). This is often a reasonable solution, but it will depend on the language whether this is reasonable solution or not. Note that for English, in spite of its somewhat poor relationship between orthography and pronunciation, this is reasonable, so don't exclude this as a possibility without trying it.

Letter selection will find the subset of the nice utterances that has the best letter coverage. It is a greedy algorithm, but this is usually sufficient. This process also takes a number to define the number of utterances it is looking for. The process will be applied multiple times to the remaining data until that number is reached. If there isn't enough data to select from it might loop for ever. By default it looks for 1000 utterance, which is not unreasonable for a unit selection voices, 500 is probably sufficient for a CLUSTERGEN voice. But, as they say, your mileage may vary.

    $FESTVOXDIR/src/promptselect/make_nice_prompts select_letter_n 

If you do have a pronunciation lexicon for you language, you can also do select based on segments rather than letters. We have not done exhaustive comparisons of how valuable segment selection is over letter selection. It is clear that although probably important, it is probably less important that selecting a good voice talent, or recording the prompts in a high quality manner. Two stages are required for segment selection. The first stage is to render the nice prompts from words to segments

    $FESTVOXDIR/src/promptselect/make_nice_prompts synth_seg

Then greedily select the utterances with the best di-phone coverage.

    $FESTVOXDIR/src/promptselect/make_nice_prompts select_seg

We do not yet support select_seg_n.

After selection the nice prompts will me in Look at it. Do not expect it to be perfect. I have never done this for a new language, without having to do it multiple times until I get something reasonable. Even once you have the result, it is worth while checking each utterance and correcting and/or rejecting it for other reasons, such as ungrammatical, hard to read, ambiguous words etc. Be bold and get rid of weird sentences, it will save you trouble later. The selection process is deliberately designed to have redundancy as speech is a variable medium and we can never be sure what exact the voice talent will say, or how the unit selection process will select the units from the database.

It is wise to first go through every sentence and attempt to record it and at that time decide if the sentence is actually a reasonable utterance to include in the prompt set for that language.

The final stage extracts the vocabulary of the selected prompt set. You can use this vocab list to start building your pronunciation lexicon as you will need that to build your speech databases (unless you are using an orthography based selection technique).

    $FESTVOXDIR/src/promptselect/make_nice_prompts find_vocab

You can also do the whole process with the command

    $FESTVOXDIR/src/promptselect/make_nice_prompts do_all TEXT0 TEXT1 ...

As really the find_nice stage takes up about 98% of processing time, redoing the other parts each time isn't unreasonable.

Non-Latin-script languages

For non-Latin-script languages there are options that seem to work well, if the language has spaces between words. We have used this quite extensively for UTF-8 encoded languages (Arabic and Hindi). For these language use

    $FESTVOXDIR/src/promptselect/make_nice_prompts select_seg
    $FESTVOXDIR/src/promptselect/make_nice_prompts find_freq_asis TEXT0 TEXT1 ...
    $FESTVOXDIR/src/promptselect/make_nice_prompts make_freq_lex
    $FESTVOXDIR/src/promptselect/make_nice_prompts find_nice_asis TEXT0 TEXT1 ...
    $FESTVOXDIR/src/promptselect/make_nice_prompts select_letter_n
    $FESTVOXDIR/src/promptselect/make_nice_prompts find_vocab_asis

You can also do the whole process with the command

    $FESTVOXDIR/src/promptselect/make_nice_prompts do_all_asis TEXT0 TEXT1 ...

For languages that do not have spaces between the words (Chinese, Japanese, Thai etc), the above techniques will not work. We have used the above techniques for Chinese, by first segmenting the data into words.