About CRBLP Speech Corpora

CRBLP speech corpora contains three speech corpora such as read speech corpus, diphone corpus and speech corpus for acoustic analysis which is suitable for uses in speech technology maintained by the http://crblp.bracu.ac.bd at BRAC University. This is specially required for speech processing applications such as Text to Speech, Speech Recognition and other speech related applications.

CRBLP read speech corpus contains high-quality recordings of a professional speaker’s voice. The aim of this project is to label the audio and text with time-aligned phonetic data that can be use to analyze the speech.
A text corpus of wide range of domains were selected that contains 1,06,860 tokens. Different domains include weekly magazine, novel, blog, legal text, a small part of constitution of Bangladesh, history, and different types of news. A professional speaker was selected for the voice and recorded in a professional voice recording studio. Recording of speech was completed in July 2008. After cleaning the recorded data, labeling was done on sentence level. This speech corpus can be used to develop acoustic models for speech recognition, to analyze the intonation pattern, and to develop a TTS by unit selection technique. This speech corpus contains ~10K sentences and ~18K unique tokens.

Diphone speech corpus contains 4355 sentences, which are typically nonsense sentences. These sentences were formed by combining the nonsense words with 4355 diphones. Diphone is a combination of two phones. In general, the number of diphone in a language is the square of the number of phones. Since Bangla language consists of 65 phones, so the number of diphones are (65X65) 4225. In addition, silence to phones are (1X65) 65 and phones to silence are (65X1) 65. So the total number of diphones are 4335. A diphone was inserted between two nonsense words to form a nensence sentence. In this way 4335 nonsence sentences were designed using each diphone in diphone speech corpus.

Speech corpus for acoustic analysis is formed to determine the number of phoneme available in Bangla. All the possible combination of phones was collected from different sources and then different patterns were selected for recording, which was then used for analysis. This resource can be used for acoustic analysis of Bangla phoneme.

The CRBLP does not guarantee the accuracy of this corpus, nor its suitablity for any specific purpose. In fact, we expect a number of errors and inconsistencies to remain in the corpus. We welcome input from users: Please send email to CRBLP (crblp@bracu.ac.bd).

The CRBLP speech corpora, is copyrighted by CRBLP at BRAC University under, Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported License. If you make use of or redistribute this material we request that you acknowledge its origin in your descriptions.

If you make any change, we would like the additions and corrections sent to us (crblp@bracu.ac.bd) for consideration in a subsequent version. All correction will be approved by the current maintainer at CRBLP.