uk-pods - speech datasets of Ukrainian podcasts.

Preparation

1. Clone the dataset repository and extract the content of clips.tar.gz archive.

git clone https://huggingface.co/datasets/taras-sereda/uk-pods
cd uk-pods && tar -zxvf clips.tar.gz

2. To use these manifests for training/inference with NeMo [1] modify audio_filepath to absolute locations of audio files extracted in previous step.

# data_root=<clonned_repo_dir> # /home/ubuntu/uk-pods
data_root=$(realpath .)
sed -i -e "s|\"audio_filepath\":\"|\"audio_filepath\":\"${data_root}\/|g" pods_train.json
sed -i -e "s|\"audio_filepath\":\"|\"audio_filepath\":\"${data_root}\/|g" pods_test.json

Usage

1. Install NeMo toolkit

pip install nemo_toolkit['all']

2. Run inference with uk-pods-conformer [2] on all files from pods_test.json manifest:

import json
from nemo.collections.asr.models import EncDecCTCModelBPE
asr_model = EncDecCTCModelBPE.from_pretrained("taras-sereda/uk-pods-conformer")

with open('pods_test.json') as fd:
    audio_paths = []
    for line in fd:
        audio_paths.append(json.loads(line)['audio_filepath'])

transcripts = asr_model.transcribe(audio_paths)

Dataset statistics

Number of wav files: 34231
Total duration: 51.066 hours
MIN duration: 1.020 sec
MAX duration: 19.999 sec
MEAN duration: 5.370 sec
MEDIAN duration: 4.640 sec

References

• [1] NVIDIA NeMo Toolkit
• [2] uk-pods-coformer ASR mode