MuST-C-clean

This is the repo for paper "On the Impact of Noises in Crowd-Sourced Data for Speech Translation" in IWSLT 2022.

This detector is adapted from code in https://pytorch.org/tutorials/intermediate/forced_alignment_with_torchaudio_tutorial.html#sphx-glr-intermediate-forced-alignment-with-torchaudio-tutorial-py.

Prepare Environment

conda create python=3.8 -n must-c-clean
conda activate must-c-clean

conda install -y pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
conda install -y tqdm pandas 
conda install -y spacy -c conda-forge
python -m spacy download en_core_web_trf

pip install editdistance num2words pyyaml

Run Detection

You can run the detection as follows:

python detect.py \
    --device {cpu/cuda} \
    --mustc-root {your must-c root directory} \
    --tgt-lang {de/other languages} \
    --split {train/dev/tst-COMMON/tst-HE}

The results will be saved in results/{split}. The tsv file mismatch.tsv contains the description of the detected audio-transcript mismatch cases. The html file mismatch.html allows you to listen to the speech and compare it with the given transcript.