Dataset Card fhnw_swiss_parliament

Dataset Summary

Swiss Parliaments Corpus, an Automatically Aligned Swiss German Speech to Standard German Text Corpus (Michel Plüss et. al.)

The inputs are a Swiss German audio recording of arbitrary length, e.g. in FLAC format, and the corresponding manual Standard German text transcript. The audio file is transcribed with Amazon Transcribe. The Amazon transcript is then globally aligned to the manual transcript using Biopython’s (Cock et al., 2009) PairwiseAligner. The manual transcript is split into sentences using spaCy (Honnibal and Montani, 2017). Each of these sentences is mapped to a start and end time in the recording via the global alignment and the per-word start and end times provided by Amazon Transcribe.

Languages

Swiss German

Dataset Structure

Data Instances

A typical tuple contains the following fields:

{'accent': '', 'age': '', 'client_id': '82', 'down_votes': '', 'gender': '',
'path': '/clips/b842e6ab-90eb-4115-bc17-a3dc49abc619.flac',
'iou_estimate'; '0.9086439204449309', 'sentence': 'Wir kommen zur Abstimmung.', 'up_votes': ''}

Data Fields

client_id: Speaker id
path: Swiss German audio clip filename
sentence: Standard German sentence text
iou_estimate: Estimated IoU (see paper)
up_votes, down_votes, age, gender, accent: Not available

Data Splits

The dataset has been subdivided into:

train_all (full train split without filtering on IoU estimate)
Number of utterances: 147370
Hours of audio: 292.56
Number of speakers: 198
Mean IoU estimate: 0.8525

train_0.7 (all sentences from train_all with an estimated IoU of > 0.7)
Number of utterances: 128779
Hours of audio: 255.74
Number of speakers: 195
Mean IoU estimate: 0.8949

train_0.9 (all sentences from train_all with an estimated IoU of > 0.9)
Number of utterances: 90324
Hours of audio: 176.43
Number of speakers: 194
Mean IoU estimate: 0.9216

test
Number of utterances: 3332
Hours of audio: 6.02
Number of speakers: 26
Mean IoU estimate: 0.9232

Instructions

Download dataset from URL https://drive.switch.ch/index.php/s/cOLa8sp2rKpf4s0
Extract .tar file using tar -xzvf <path/to/downloaded/file>
Contains files: README.md, test.tsv, train_0.7.tsv, train_0.9.tsv, train_all.tsv
Audio samples in folder /clips are sampled at 16kHz Sample length varies however and needs to be adjusted accordingly
Fhnw_swiss_parliament.py allows splits ['train_all.tsv','train_0.7.tsv','train_0.9.tsv','test','train','validation','other']

from datasets import load_dataset
fhnw_swiss_parliament_train = load_dataset("Yves/fhnw_swiss_parliament", data_dir="./SwissParliament", split="train_0.9")
fhnw_swiss_parliament_validation = load_dataset("Yves/fhnw_swiss_parliament", data_dir="./SwissParliament", split="test")

Additional Information

@inproceedings{
  author = {Michel Pluss, Lukas Neukom, Manfred Vogel},
  title = {Swiss Parliaments Corpus, an Automatically Aligned Swiss German Speech to Standard German Text Corpus},
  pages = {1-5},
  year = 2020
}