Metawave

Metawave is a command line tool for generating reader meta information from TTS datasets. This tool can be used for analyzing and comparing reader in multi-reader TTS datasets. Metawave can generate a list of outliers in the dataset that are potentially better left out in TTS training. Metawave only supports .wav for the audio but should work with any regular file type for text tokens.

Installing

Metawave is a python 3.6 application and to run it, all dependencies need to be installed first via pip install -r requirements.txt. Even better is to install metawave as a package. To use metwave as a package install it via pip by cloning this repository, and inside the base metawave directory do pip install .

Usage

All running of the app is through metawave.py. To get more information about how to run metwave you can always use -h to get information about how to run. Metwave has 6 available commands:n

1. metawave --run : For gathering the initial meta-information and optionally generates a summary as well. (Requires an index file)
2. metawave --custom_run : Can be used for unsupported datasets. (Requires an index file)
3. metawave --summary : Can only be called after doing a --run. This summary generates information about each speaker and a list of outliers amongst other things.
4. metawave --check : A handy tool to get quick information ona single <text, audio> pair.
5. metawave --outliers : Generates an outlier index for a given dataset meta file.
6. metawave --gen_index : For generating a line index file similar to the one in the google dataset for example.
7. metawave --search: For searching for terms in text tokens.

Each mode has some required parameters which can be listed via metawave --<mode> -h

Generating an index file

If a dataset does not have an index file, one has to be created. If the reader id is encoded in the file-name, this can be specified via --name_reg. For example if the data has the following setup:

• text-001-001.wav
• text-001-002.wav
• ...
• text-002-001.wav

where the first 3 numbers are an id for the reader and the last 3 are a file id for the <text, audio> pair, we can supply --name_reg .*-r-i where r stands for reader and i for file id. Then the reader information will be saved. If nothing is supplied to the name regular expression, we assume that the full name of each <text,audio> pair is the same and the full name will be used as an id for the pair. Note that the regular expression is dependent on the file name format of the text tokens. Currently this functionality only supports those regular variables , .*, r, i and any delimiters. Other examples could be:

• --name_reg r_.*_i matches e.g. <reader_id>_token_<file_id>
• --name_reg .*-i matches e.g. token_<file_id> where the <reader_id> is not encoded in the file name.

Also note that the file extension, i.e. the .txt in token.txt, should not be included.

Custom datasets

Since the directory structure of each dataset can be different, the only supported datasets are TTS_icelandic_Google_m and ivona_speech_data. However, for unsupported datasets, you can run metawave by supplying --wav_dir, --text_dir and --index_path to the --custom_run command. For this to work, a line index file has to be supplied and at least --wav_ind and --txt_ind which are the indices of audio filenames and text token filenames in each line respectively. So if a line in an index file is <text_fname> \t <audio_fname> \t <reader_id> the suggested arguments would be --txt_ind 0 --wav_ind 1 --reader_ind 2.

Using Ivona To use Ivona as a known dataset, an index file, line_index.tsv has to be generated and available in the root directory of the dataset. Each line has <text_fname> \t <audio_fname> \t <reader_id>. In this case, the reader id references the 3 different sets in Ivona. Otherwise an index file can be generated anyway is needed and run using custom_run