Multiple I/O

Most of the time, conversion pipelines will only need to read from one source and output to a single target. However, there can be cases where datasets of different types need merging (multiple inputs) or datasets of different types need to generated for different frameworks (multiple outputs).

To cater for these scenarios, the following two meta plugins are available:

• from-multi - reads from one or more sources using the specified readers
• to-multi - forwards the incoming data to one or more writers

There is one restriction, each of the base reader/writer must be from the same data domain.

Multiple inputs

The following command reads a dataset in Festvox and ADAMS format, with the combined output being saved in CommonVoice format:

adc-convert \
  -l INFO \
  from-multi \
    -l INFO \
    -t sp \
    -r "from-festvox-sp -l INFO -i {CWD}/input/*.txt" \
       "from-adams-sp -l INFO -i {CWD}/input/*.report -t transcript" \
  to-commonvoice-sp \
    -l INFO \
    -o "{CWD}/output"

Multiple outputs

Below, the source data is in ADAMS speech format and will be converted to Festvox and CommonVoice format:

adc-convert \
  -l INFO \
  from-adams-sp \
    -l INFO \
    -i "{CWD}/input/*.report" \
  to-multi \
    -l INFO \
    -t sp \
    -w "to-festvox-sp -l INFO -o {CWD}/output/festvox" \
       "to-commonvoice-sp -l INFO -o {CWD}/output/commonvoice"

Sub-pipelines

With the tee meta-filter, it is possible to filter the audio files coming through with a separate sub-pipeline. E.g., converting the incoming data into multiple output formats with their own preprocessing.

The following command loads the Festvox speech data and saves them in ADAMS and split ADAMS format (after trimming silences) in one command:

adc-convert \
  -l INFO \
  from-festvox-sp \
    -l INFO \
    -i "./festvox/*.txt" \
  tee \
    -f "to-adams-sp -o ./adams-tee/ -t transcript" \
  tee \
    -f "trim-silence to-adams-sp -o ./adams-split-tee/ -t transcript --split_names train val test --split_ratios 70 15 15"