Datasets:

Modalities:
Audio
Text
Formats:
parquet
ArXiv:
Libraries:
Datasets
Dask
License:

Text data in Malayalam is corrupted with consonants+virama instead of chillus

#2
by kavyamanohar - opened

It looks like the text data source was probably corrupted, or there was an automated script that removed all zwj from the dataset.

Errors and corrections highlighted from the first few text field of the dataset. Since chillus are pretty common in Malayalam script and their corruption not only affects readabiliy, sometimes meaning too. Eg: അവൻ -> അവന് (He->Him)

fname Original Corrected
844424933086206-462-f.m4a വാഹനത്തിന്റെ സുരക്ഷ ഉറപ്പാക്കുന്നതിലും കംഫര്ട്ടബിള് ആയിട്ടുള്ള ഡ്രൈവിങ്ങ് ഒരുക്കുന്നതിലും ഇത്തരം ഫീച്ചറുകള് വലിയ പങ്കുവഹിക്കുന്നുണ്ട് വാഹനത്തിന്റെ സുരക്ഷ ഉറപ്പാക്കുന്നതിലും കംഫട്ടബി ആയിട്ടുള്ള ഡ്രൈവിങ്ങ് ഒരുക്കുന്നതിലും ഇത്തരം ഫീച്ചറുക വലിയ പങ്കുവഹിക്കുന്നുണ്ട്
844424933049315-462-f.m4a നേപ്പാളിന്റെയും ചൈനയുടെയും ആദരവ് ഏറ്റുവാങ്ങാന് കഴിഞ്ഞ സംന്യാസിയാണ് ജയേന്ദ്ര സരസ്വതി സ്വാമികള് നേപ്പാളിന്റെയും ചൈനയുടെയും ആദരവ് ഏറ്റുവാങ്ങാ കഴിഞ്ഞ സംന്യാസിയാണ് ജയേന്ദ്ര സരസ്വതി സ്വാമിക
844424932990980-462-f.m4a ഇതിലൂടെ കാലതാമസം കൂടാതെ കയറ്റുമതിക്കാര്ക്ക് തങ്ങളുടെ ഉല്പ്പന്നം വേണ്ട സ്ഥലത്ത് എത്തിക്കാന് സാധിക്കും ഇതിലൂടെ കാലതാമസം കൂടാതെ കയറ്റുമതിക്കാർക്ക് തങ്ങളുടെ ഉല്പ്പന്നം വേണ്ട സ്ഥലത്ത് എത്തിക്കാ സാധിക്കും
844424931260721-462-f.m4a ഇസ്ലാമിക സാമ്രാജ്യം സ്വപ്നം കാണുന്ന എര്ദോഗനുമായുള്ള ബന്ധം അപകടകരമാണെന്ന് വിവിധ മുസ്ലിം സംഘടനകള് ചൂണ്ടിക്കാട്ടുന്നു ഇസ്ലാമിക സാമ്രാജ്യം സ്വപ്നം കാണുന്ന എദോഗനുമായുള്ള ബന്ധം അപകടകരമാണെന്ന് വിവിധ മുസ്ലിം സംഘടനക ചൂണ്ടിക്കാട്ടുന്നു

Automated correction seems impossible. Since Kathbath is a popular benchmark for Indic ASR, this data corruption has larger impacts.

Are there any thoughts on correcting the dataset?

Sign up or log in to comment