Train-clean-100 dataset

LibriSpeech ：

It is a reading phonetic corpus , be based on LibriVox Public domain audiobooks . Its purpose is to realize automatic speech recognition (ASR) System training and testing .

The corpus is divided into several parts , So that users can selectively download its subset according to their own needs . The name carries “clean” A subset of is considered more than other audio and American English accents “ clean ”（ At least on average ）. This classification is obtained using very rough automated means , It should not be considered completely reliable . Subsets are disjoint , That is, each speaker's audio is assigned to a subset .

The structure of the corpus is as follows ：

* dev-clean, test-clean - contain “ clean ” Voice development and test set .

* train-clean-100 - Training set , about 100 Hours of “ clean ” voice
* train-clean-360 - Training set , about 360 Hours of “ clean ” voice

* dev-other, test-other - Development and test sets , Voice is automatically selected to be more “ challenging ” The identification of

* train-other-500 - about 500 Hour training set , Contains are not classified as “ clean ” The voice of

* intro - subset  Only some readers LibriVox Introduce a subset of disclaimers

* mp3  - The original on which the corpus is based MP3 Encode audio

* texts - The text corresponding to the audio in the corpus

1. train-clean-100 The architecture of the dataset ：

among ：

【19】： Speaker's ID

【198/227】： The chapter of the article that the speaker reads

【*.trans.txt】： Contains text from the corresponding chapter , and FLAC The document is the voice of each chapter .