Speech Datasets Collection-语音数据集汇总

来源：https://github.com/RevoSpeechTech/speech-datasets-collection

openslr 下载：

1）修改为国内地址

例如 aishell，默认的run.sh里写的是www.openslr.org/resources/33，需要改为国内站点，http://openslr.magicdatatech.com/resources/33。

其他目录可以看： http://openslr.magicdatatech.com/resources.php

在使用 wget 下载文件时，如果遇到下载速度慢的问题，可以通过以下几种方法加速下载：

1. 使用多个连接

wget 默认只使用单个连接进行下载，但是你可以使用 aria2 这种工具，它支持多线程下载，显著加速下载速度。aria2 可以通过以下命令安装：

sudo apt install aria2  # For Ubuntu/Debian
brew install aria2  # For macOS

然后你可以使用 aria2 下载文件：

aria2c -x 16 -s 16 <URL>

-x 16 表示使用 16 个连接来下载文件，-s 16 表示将下载源分为 16 个部分。

2. 使用 --limit-rate 限制下载速度

虽然这并不会直接加速下载，但如果下载的速度不稳定，设置一个合理的速率限制可以避免带宽波动影响速度。在命令中加上 --limit-rate 参数：

wget --limit-rate=1m <URL>

这将限制下载速度为每秒 1 MB。

3. 启用断点续传

如果下载过程中断，可以使用 -c 或 --continue 参数来启用断点续传，从中断的地方继续下载：

wget -c <URL>

This is a curated list of open speech datasets for speech-related research (mainly for Automatic Speech Recognition).

Over 110 speech datasets are collected in this repository, and more than 70 datasets can be downloaded directly without further application or registration.

Notice:

1. This repository does not show corresponding License of each dataset. Basically it’s OK to use these datasets for research purpose only. Please make sure the License is suitable before using for commercial purpose.
2. Some small-scale speech corpora are not shown here for concision.

1. Data Overview

Dataset Acquisition	Sup/Unsup	All Languages (Hours)	Mandarin (Hours)	English (Hours)
download directly	supervised	199k +	2110 +	34k +
download directly	unsupervised	530k +	1360 +	68k +
download directly	total	729k +	3470 +	102k +
need application	supervised	53k +	16740 +	50k +
need application	unsupervised	60k +	12400 +	57k +
need application	total	113k +	29140 +	107k +
total	supervised	252k +	18850 +	84k +
total	unsupervised	590k +	13760 +	125k +
total	total	842k +	32610 +	209k +

• Mandarin here includes Mandarin-English CS corpora.
• Sup means supervised speech corpus with high-quality transcription.
• Unsup means unsupervised or weakly-supervised speech corpus.

2. List of ASR corpora

a. datasets can be downloaded directly

id	Name	Language	Type/Domain	Paper Link	Data Link	Size (Hours)
1	Librispeech	English	Reading	[paper]	[dataset]	960
2	TED_LIUM v1	English	Talks	[paper]	[dataset]	118
3	TED_LIUM v2	English	Talks	[paper]	[dataset]	207
4	TED_LIUM v3	English	Talks	[paper]	[dataset]	452
5	MLS	Multilingual	Reading	[paper]	[dataset]	50k +
6	thchs30	Mandarin	Reading	[paper]	[dataset]	35
7	ST-CMDS	Mandarin	Commands	–	[dataset]	100
8	aishell	Mandarin	Recording	[paper]	[dataset]	178
9	aishell-3	Mandarin	Recording	[paper]	[dataset]	85
10	aishell-4	Mandarin	Meeting	[paper]	[dataset]	120
11	aishell-eval	Mandarin	Misc	–	[dataset]	80 +
12	Primewords	Mandarin	Recording	–	[dataset]	100
13	aidatatang_200zh	Mandarin	Record	–	[dataset]	200
14	MagicData	Mandarin	Recording	–	[dataset]	755
15	MagicData-RAMC	Mandarin	Conversational	[paper]	[dataset]	180
16	Heavy Accent Corpus	Mandarin	Conversational	–	[dataset]	58 +
17	AliMeeting	Mandarin	Meeting	[paper]	[dataset]	120
18	CN-Celeb	Mandarin	Misc	[paper]	[dataset]	unsup(274)
19	CN-Celeb2	Mandarin	Misc	[paper]	[dataset]	unsup(1090)
20	The People’s Speech	English	Misc	[paper]	[dataset]	30k +
21	Multilingual TEDx	Multilingual	Talks	[paper]	[dataset]	760 +
22	VoxPopuli	Multilingual	Misc	[paper]	[dataset]	sup(1.8k) unsup(400k)
23	Libri-Light	English	Reading	[paper]	[dataset]	unsup(60k)
24	Common Voice (Multilingual)	Multilingual	Recording	[paper]	[dataset]	sup(15k) unsup(5k)
25	Common Voice (English)	English	Recording	[paper]	[dataset]	sup(2200) unsup(700)
26	JTubeSpeech	Japanese	Misc	[paper]	[dataset]	1300
27	ai4bharat NPTEL2020	English(Indian)	Lectures	–	[dataset]	weaksup(15.7k)
28	open_stt	Russian	Misc	–	[dataset]	20k +
29	ASCEND	Mandarin-English CS	Conversational	[paper]	[dataset]	10 +
30	Crowd-Sourced Speech	Multilingual	Recording	[paper]	[dataset]	1200 +
31	Spoken Wikipedia	Multilingual	Recording	[paper]	[dataset]	1000 +
32	MuST-C	Multilingual	Talks	[paper]	[dataset]	6000 +
33	M-AILABS	Multilingual	Reading	–	[dataset]	1000
34	CMU Wilderness	Multilingual	Misc	[paper]	[dataset]	unsup(14k)
35	Gram_Vaani	Hindi	Recording	[paper] [code]	[dataset]	sup(100) unsup(1k)
36	VoxLingua107	Multilingual	Misc	[paper]	[dataset]	unsup(6600 +)
37	Kazakh Corpus	Kazakh	Recording	[paper] [code]	[dataset]	335
38	Voxforge	English	Recording	–	[dataset]	130
39	Tatoeba	English	Recording	–	[dataset]	200
40	IndicWav2Vec	Multilingual	Misc	[paper]	[dataset]	unsup(17k +)
41	VoxCeleb	English	Misc	[paper]	[dataset]	unsup(352)
42	VoxCeleb2	English	Misc	[paper]	[dataset]	unsup(2442)
43	RuLibrispeech	Russian	Read	–	[dataset]	98
44	MediaSpeech	Multilingual	Misc	[paper]	[dataset]	40
45	MUCS 2021 task1	Multilingual	Misc	–	[dataset]	300
46	MUCS 2021 task2	Multilingual	Misc	–	[dataset]	150
47	nicolingua-west-african	Multilingual	Misc	[paper]	[dataset]	140 +
48	Samromur 21.05	Samromur	Misc	[code]	[dataset] [dataset][dataset]	145
49	Puebla-Nahuatl	Puebla-Nahuatl	Misc	[paper]	[dataset]	150 +
50	Golos	Russian	Misc	[paper]	[dataset]	1240
51	ParlaSpeech-HR	Croatian	Parliament	[paper]	[dataset]	1816
52	Lyon Corpus	French	Recording	[paper]	[dataset]	185
53	Providence Corpus	English	Recording	[paper]	[dataset]	364
54	CLARIN Spoken Corpora	Czech	Recording	–	[dataset]	1120 +
55	Czech Parliament Plenary	Czech	Recording	–	[dataset]	444
56	(Youtube) Regional American Corpus	English (Accented)	Misc	[paper]	[dataset]	29k +
57	NISP Dataset	Multilingual	Recording	[paper]	[dataset]	56 +
58	Regional African American	English (Accented)	Recording	[paper]	[dataset]	130 +
59	Indonesian Unsup	Indonesian	Misc	–	[dataset]	unsup (3000+)
60	Librivox-Spanish	Spanish	Recording	–	[dataset]	120
61	AVSpeech	English	Audio-Visual	[paper]	[dataset]	unsup(4700)
62	CMLR	Mandarin	Audio-Visual	[paper]	[dataset]	100 +
63	Speech Accent Archive	English	Accented	[paper]	[dataset]	TBC
64	BibleTTS	Multilingual	TTS	[paper]	[dataset]	86
65	NST-Norwegian	Norwegian	Recording	–	[dataset]	540
66	NST-Danish	Danish	Recording	–	[dataset]	500 +
67	NST-Swedish	Swedish	Recording	–	[dataset]	300 +
68	NPSC	Norwegian	Parliament	[paper]	[dataset]	140
69	CI-AVSR	Cantonese	Audio-Visual	[paper]	[dataset]	8 +
70	Aalto Finnish Parliament	Finnish	Parliament	[paper]	[dataset]	3100 +
71	UserLibri	English	Reading	[paper]	[dataset]	–
72	Ukrainian Speech	Ukrainian	Misc	–	[dataset]	1300+
73	UCLA-ASR-corpus	Multilingual	Misc	–	[dataset]	unsup(15k) sup(9k)
74	ReazonSpeech	Japanese	Misc	[paper] [code]	[dataset]	15k
75	Bundestag	German	Debate	[paper]	[dataset]	sup(610) unsup(1038)

b. datasets can be downloaded after application

id	Name	Language	Type/Domain	Paper Link	Data Link	Size (Hours)
1	Fisher	English	Conversational	[paper]	[dataset]	2000
2	WenetSpeech	Mandarin	Misc	[paper]	[dataset]	sup(10k) weaksup(2.4k) unsup(10k)
3	aishell-2	Mandarin	Recording	[paper]	[dataset]	1000
4	aidatatang_1505zh	Mandarin	Recording	–	[dataset]	1505
5	SLT 2021 CSRC	Mandarin	Misc	[paper]	[dataset]	400
6	GigaSpeech	English	Misc	[paper]	[dataset]	sup(10k) unsup(23k)
7	SPGISpeech	English	Misc	[paper]	[dataset]	5000
8	AESRC 2020	English (accented)	Misc	[paper]	[dataset]	160
9	LaboroTVSpeech	Japanese	Misc	[paper]	[dataset]	2000 +
10	TAL_CSASR	Mandarin-English CS	Lectures	–	[dataset]	587
11	ASRU 2019 ASR	Mandarin-English CS	Reading	–	[dataset]	700 +
12	SEAME	Mandarin-English CS	Recording	[paper]	[dataset]	196
13	Fearless Steps	English	Misc	–	[dataset]	unsup(19k)
14	FTSpeech	Danish	Meeting	[paper]	[dataset]	1800 +
15	KeSpeech	Mandarin	Recording	[paper]	[dataset]	1542
16	KsponSpeech	Korean	Conversational	[paper]	[dataset]	969
17	RVTE database	Spanish	TV	[paper]	[dataset]	800 +
18	DiDiSpeech	Mandarin	Recording	[paper]	[dataset]	800
19	Babel	Multilingual	Telephone	[paper]	[dataset]	1000 +
20	National Speech Corpus	English (Singapore)	Misc	[paper]	[dataset]	3000 +
21	MyST Children’s Speech	English	Recording	–	[dataset]	393
22	L2-ARCTIC	L2 English	Recording	[paper]	[dataset]	20 +
23	JSpeech	Multilingual	Recording	[paper]	[dataset]	1332 +
24	LRS2-BBC	English	Audio-Visual	[paper]	[dataset]	220 +
25	LRS3-TED	English	Audio-Visual	[paper]	[dataset]	470 +
26	LRS3-Lang	Multilingual	Audio-Visual	–	[dataset]	1300 +
27	QASR	Arabic	Dialects	[paper]	[dataset]	2000 +
28	ADI (MGB-5)	Arabic	Dialects	[paper]	[dataset]	unsup (3000 +)
29	MGB-2	Arabic	TV	[paper]	[dataset]	1200 +
30	3MASSIV	Multilingual	Audio-Visual	[paper]	[dataset]	sup(310) unsup(600)
31	MDCC	Cantonese	Misc	[paper]	[dataset]	73 +
32	Lahjoita Puhetta	Finnish	Misc	[paper]	[dataset]	sup(1600) unsup(2000)
33	SDS-200	Swiss German	Dialects	[paper]	[dataset]	200
34	Modality Corpus	Multilingual	Audio-Visual	[paper]	[dataset]	30 +
35	Hindi-Tamil-English	Multilingual	Misc	–	[dataset]	690
36	English-Vietnamese Corpus	English, Vietnamese	Misc	[paper]	[dataset]	500+
37	OLKAVS	Korean	Audio-Visual	[paper] [code]	[dataset]	1150

3. References

• https://github.com/coqui-ai/open-speech-corpora
• https://openslr.org/resources.php