README.md

metadata

language:
  - ur
task_categories:
  - automatic-speech-recognition
dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: transcription
      dtype: string
  splits:
    - name: train
      num_bytes: 108262823.43502825
      num_examples: 566
    - name: test
      num_bytes: 27161338.564971752
      num_examples: 142
  download_size: 135363971
  dataset_size: 135424162
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

The Urdu Phonetically Rich Speech Corpus consists of 70 minutes of transcribed read speech consisting of 708 greedily created sentences representing all phonemic and triphonemic combinations in Urdu (based on an 18 million word corpus of Urdu news articles). It consists of 10,101 tokens with 5,656 unique words. In addition to providing phonetic cover for Urdu, the corpus is also phonemically balanced. It also provides triphonemic cover however it is not completely balanced for triphonemes. It contains 60 unique phones and 42,289 phone occurrences. The sentences contained in this corpus are all manually created by trained linguists following a greedy approach to accommodate the words (which were selected using a set cover algorithm) and to prevent additional words as much as possible. Therefore, while correct grammatically, there are some instances where the choice of words in the sentences is unusual.

Further information and download instructions can be found at https://www.c-salt.org/downloads/prus

Copyright (c) 2017 by Center for Speech and Language Technologies (CSaLT), Information Technology University of the Punjab, Lahore, Pakistan. Your use of the CSaLT Phonetically Rich Urdu Speech Corpus is subject to our Creative Commons License (https://creativecommons.org/licenses/by/4.0/), which lets you distribute, remix, tweak, and build upon our work, even commercially, as long as you credit us for the original creation. You are required to cite the "Center for Speech and Language Technologies (CSaLT)" and the following two publications:

1. Agha Ali Raza, Sarmad Hussain, Huda Sarfraz, Inam Ullah, Zahid Sarfraz, An ASR System for Spontaneous Urdu Speech, Oriental COCOSDA 2010 conference, Nov. 24-25, 2010, Katmandu, Nepal.

2. Agha Ali Raza, Sarmad Hussain, Huda Sarfraz, Inam Ullah, Zahid Sarfraz, Design and development of phonetically rich Urdu speech corpus, Proceedings of O-COCOSDA'09 and IEEE Xplore; O-COCOSDA'09, 10-13 Aug 2009, School of Information Science and Engineering of Xinjiang University, Urunqi, China (URL: http://o-cocosda2009.xju.edu.cn).