File size: 3,017 Bytes
c387b85
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
INTRODUCTION
============
Multilingual LibriSpeech (MLS) dataset is a large multilingual corpus 
suitable for speech research. The dataset is derived from read audiobooks 
from LibriVox and consists of 8 languages - English, German, Dutch, French, 
Spanish, Italian, Portuguese and Polish. 


STATISTICS
==========
------------------------------------------------
Language   |  train(hrs) | dev(hrs) | test(hrs)
------------------------------------------------
English	   |  44659.74   |  15.75   |  15.55
German	   |   1966.51	 |  14.28   |  14.29
Dutch	   |   1554.24	 |  12.76   |  12.76
French	   |   1076.58	 |  10.07   |  10.07
Spanish	   |    917.68	 |   9.99   |  10.00
Italian	   |    247.38	 |   5.18   |   5.27
Portuguese |	160.96	 |   3.64   |   3.74
Polish	   |    103.65	 |   2.08   |   2.14
------------------------------------------------


DIRECTORY STRUCTURE
===================

A typical top-level directory layout

.
β”œβ”€β”€ metainfo.txt                                # Info about the LibriVox data used in the dataset (one row per chapter)
β”œβ”€β”€ dev                                         # Dev (validation) subset
β”‚Β Β  β”œβ”€β”€ segments.txt
β”‚Β Β  β”œβ”€β”€ transcripts.txt
β”‚Β Β  └── audio
β”‚       └── ...
β”œβ”€β”€ test                                        # Test subset
β”‚Β Β  β”œβ”€β”€ segments.txt
β”‚Β Β  β”œβ”€β”€ transcripts.txt
β”‚Β Β  └── audio
β”‚       └── ...
β”œβ”€β”€ train                                    	# Train subset 
β”‚Β Β  β”œβ”€β”€ segments.txt                            # LibriVox audio links and segment start/end timestamps (in sec) for each audio file
β”‚Β Β  β”œβ”€β”€ transcripts.txt                         # Transcripts file with audio file name as key
β”‚Β Β  β”œβ”€β”€ audio 
β”‚   β”‚   β”œβ”€β”€ 10148   	 			# LibriVox speaker ID
β”‚   β”‚   β”‚   β”œβ”€β”€ 10119 				# LibriVox book ID
β”‚   β”‚   β”‚   β”‚   β”œβ”€β”€ 10148_10119_000000.flac	# 16Khz audio file
β”‚   β”‚   β”‚   β”‚   β”œβ”€β”€ 10148_10119_000001.flac
β”‚   β”‚   β”‚   β”‚   └── ...
β”‚   β”‚   β”‚   └── ...       
β”‚   β”‚   └── ...
β”‚Β Β  └── limited_supervision			# Limited supervision training sets (10 min, 1 hr and 10 hr)
β”‚       β”œβ”€β”€ 1hr                               	# Data of the 1h split (made up of 6 folds of 10 min)
β”‚       β”‚   β”œβ”€β”€ 0				# first 10 min fold
β”‚       β”‚   β”‚   └── handles.txt			# List of audio handles for the split	
β”‚       β”‚   β”œβ”€β”€ ...		
β”‚       β”‚   └── 6                               # last 10 min fold
β”‚       β”‚       └── handles.txt                 
β”‚       └── 9hr					# Remaining data of the 10h split (10h=1h+9h)
β”‚           └── handles.txt
β”œβ”€β”€ LICENSE				        # License info
└── README

For more information, see the paper "MLS: A Large-Scale Multilingual Dataset for Speech Research", Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve and Ronan Collobert, INTERSPEECH 2020.