Code for checking goodness of data for STT and TTS.
$ git clone https://github.com/coqui-ai/data-checker.git
$ cd data-checker
$ docker build . -t data-checker
$ docker run data-checker python data_checks.py "/code/data/smoke_test/russian_sample_data/ru.csv" 2
.
.
.
π β Found 1 <transcript,clip> pairs in /code/data/smoke_test/russian_sample_data/ru.csv
Β· First audio file found: ru.wav of type audio/wav
Β· Checking if audio is readable...
π Found no unreadable audiofiles
Β· Reading audio duration...
π β Found a total of 0.00 hours of readable data
Β· Get transcript length...
Β· Get num feature vectors...
π Found no audio clips over 30 seconds in length
π Found no transcripts under 10 characters in length
Β· Get ratio (num_feats / transcript_len)...
π Found no offending <transcript,clip> pairs
Β· Calculating ratio (num_feats : transcript_len)...
π Found no <transcript,clip> pairs more than 2.0 standard deviations from the mean
π β¬ Saved a total of 0.00 hours of data to BEST dataset
β Removed a total of 0.00 hours (0.00% of original data)
β Removed a total of 0 samples (0.00% of original data)
β Wrote best data to /code/data/smoke_test/russian_sample_data/ru.BEST
data-checker
assumes your CSV has two columns: wav_filename
and transcript
. Note that you don't actually need to use WAV files, but the header still should be wav_filename
.
$ docker run data-checker --mount "type=bind,src=/path/to/my/local/data,dst=/mnt" python data_checks.py "/mnt/my-data.csv" 2