Uzbek Speech Corpus and Automatic Speech Recognition


Open Source Database of speech corpus and designed to create a system of automatic speech recognition in the Uzbek language and the automatic speech recognition system.

Start

  Visitors
1604

  Total number of audios
107324

  Total audio duration
118.2 hours (01.07.2021)

Audio collection process

Assalumu alaykum. You have visited the system designed to form a speech corpus for the system of automatic recognition of the Uzbek language point. Thank you for your help. Your help is very important to us.


* To use the system, you will first need to fill out a survey (in case of first visit).
* You will be asked to read the text by pressing the microphone button once and then press the microphone button again.
* Listen to the generated audio file and see if it is correct. click the " Submit " button.
* You can rewrite the audio if it is read incorrectly.
* Click the " Generate text " button to change the text provided to you!



Survey



    
  
            
  






Audio collection process (via telegram bot)


In order to make the use of the system more user-friendly, a bot with a special name UzSpeechDB_bot was created on the social network Telegram. You can go to the telegram bot page by clicking the link provided!

This project has been developed in collaboration between the Image and Speech Processing Laboratory of the Department of Computer Systems of the Tashkent University of Information Technologies named after Muhammad al-Khwarizmi and the Institute of Smart Systems and Artificial Intelligence (ISSAI) and is protected by Creative Commons Attribution 4.0 International License. The work carried out under the project is reflected in the following scientific article:


Musaev, M., Mussakhojayeva, S., Khujayorov, I., Khassanov, Y., Ochilov, M., & Varol, H. A. (2020). USC: An Open-Source Uzbek Speech Corpus and Initial Speech Recognition Experiments. arXiv preprint arXiv:2107.14419.

Technology for the development of speech recognition systems in the Uzbek language, presented in the link below https://github.com/Smart-Projects-Artificial-Intelligence/Uzbek-ASR


Automatic Speech Recognition


Below is the demo of the automatic speech recognition system built using Uzbek speech corpus. Please press the microphone button and speak immediately until the counter reaches zero. The recognized output will be displayed above the microphone button after 10 seconds. Note that some browsers do not support voice recording features !!!


* First, select the model. ( Pay attention to the indicators of WER-Word Error Rate and CER-Character Error Rate !)
* Choose the language model (LM-Language model) corresponding to the selected model.
* Press the microphone button and speak immediately until the counter reaches zero.
* The recognized output will be displayed above the microphone button after 10 seconds.
* The microphone button is pressed again to perform a new test.




Testing process


      


      






      






Statistics

Collected data statistics (01.07.2021)


Duration (hours) 100.2 10.8 7.2 118.2
# Utterances 90.012 7.321 5.211 104.544
# Words 451.1k 31.3k 30.2k 512.6k
# Unique Words 50.2k 11.2k 13.1k 74.5k
# Speakers 882 83 67 1032
The USC dataset specifications.




Duration (seconds)
(a)
Length (words)
(b)

(a) Distribution of the duration of the audios that make up the body, (b) Distribution of the number of words in the texts that make up the body.








Age and gender statistics of announcers involved in corpus formation.

Contact with us


If you want to cooperate with us and want to use the collected database, you can contact with us



Uzbek Speech Corpus and Automatic Speech Recognition

Open Source Database of speech corpus designed to create a system of automatic speech recognition in the Uzbek language and the automatic speech recognation.

TUIT, Department of Artificial Intelligence, Laboratory of Image and Speech Signal Processing

ravotcha1992@gmail.com

+998 94 651 64 51