AVMuST-TED

The first Audio-Visual Multi-lingual Speech Translation dataset.

Paper: MixSpeech: Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition

Introduction

AVMuST-TED contains over 706 hours of videos and the translation text of 4 different languages.

Target Language	Hours	Sents	Vocab	Tokens	Labeling Type
Spanish (Es)	198h	258K	95K	2.0M	Machine Translation + Manual Review
French (Fr)	185h	244K	91K	1.9M	Machine Translation + Manual Review
Italian (It)	165h	218K	95K	1.6M	Machine Translation + Manual Review
Portuguese (Pt)	158h	205K	84K	1.5M	Machine Translation + Manual Review

Acquisition of AVMuST-TED

• Transcription & Translation

  The transcription and translations in the AVMuST-TED dataset are taken directly from the high reliability translated subtitles in TED. All translations and transcriptions are performed strictly in accordance with the TED Translation Guidelines. The transcription and translations go through the following steps before publication:

  

  1. Transcription

     TED provides an original transcript for all TED and TED-Ed content. For TEDx talks, volunteers are able to utilize auto-generated transcriptions as a base, or create their own from scratch.

  2. Translation

     Subtitles are translated from the original language into the target language, using a dynamic subtitle editor. Note that TEDx transcripts must be reviewed and published before any translations can be created.

  3. Review

     Before publication, subtitles are reviewed by an experienced volunteer. Once reviewed, subtitles publish directly to the TED.com video player and YouTube for the world to enjoy.

• Audio and Visual Speech Preparation

  We follow the previous audio-video speech dataset acquisitions pipeline to extract audio speech and visual speech (i.e., talking head). The visual speech corpus is provided as face-centered video in .avi files with a resolution of 224*224 and a frame rate of 25fps. The audio speech corpus is provided as the single-track, 16-bit 16kHz .wav files.

• AV Synchronization and Active Speaker Detection

  To detect active speakers, SyncNet is further used to determine which face's lip movements match the audio, and if there is no match, the clip is rejected as being a voice-over.

Tasks on AVMuST-TED

• Visual Speech Translation (Lip Translation)

Method	M	En-Es	En-Fr	En-It	En-Pt
Cascaded	V	12.7	11.3	11.5	13.2
AV-Hubert	V	14.2	12.6	12.9	14.8
Cascaded	A(+Noise)	16.0	12.9	12.6	15.5
AV-Hubert	A(+Noise)	17.6	14.5	14.1	17.1
MixSpeech(ours)	V	18.5	15.1	14.3	17.2

• Audio Speech Translation

Src-Tgt	Method	M	-20db	-10db	0db	10db	20db	Avg.	Clean
En-Es	Cascaded	A	1.4±0.1	5.8±0.2	21.1±0.3	25.5±0.3	26.3±0.2	16.0	26.6
	AV-Hubert	A	1.5±0.2	6.7±0.2	22.3±0.4	27.7±0.2	28.6±0.3	17.6	28.9
En-Fr	Cascaded	A	1.3±0.2	4.5±0.3	16.6±0.4	20.9±0.3	21.3±0.1	12.9	21.7
	AV-Hubert	A	1.4±0.2	5.5±0.3	18.5±0.4	23.2±0.2	23.6±0.1	14.5	23.9
En-It	Cascaded	A	0.9±0.3	4.0±0.3	16.1±0.2	20.7±0.1	21.2±0.2	12.6	21.5
	AV-Hubert	A	1.0±0.2	5.1±0.5	18.3±0.3	22.7±0.2	23.6±0.2	14.1	23.8
En-Pt	Cascaded	A	1.1±0.3	5.4±0.5	20.1±0.4	24.9±0.2	26.0±0.1	15.5	26.2
	AV-Hubert	A	1.2±0.2	6.3±0.4	22.2±0.3	27.4±0.3	28.4±0.1	17.1	28.6

• Audio-Visual Speech Translation

Src-Tgt	Method	M	-20db	-10db	0db	10db	20db	Avg.	Clean
En-Es	Cascaded	AV	6.7±0.2	15.3±0.4	24.6±0.4	26.3±0.2	26.7±0.2	19.9	26.9
	AV-Hubert	AV	6.9±0.3	16.4±0.5	26.6±0.3	28.7±0.1	28.9±0.2	21.5	29.1
En-Fr	Cascaded	AV	4.6±0.1	11.4±0.5	19.4±0.3	21.5±0.2	22.0±0.2	15.8	22.3
	AV-Hubert	AV	4.9±0.2	12.1±0.3	21.6±0.4	23.7±0.3	24.3±0.1	17.3	24.6
En-It	Cascaded	AV	4.8±0.3	11.8±0.4	19.5±0.3	21.4±0.2	22.1±0.1	15.9	22.3
	AV-Hubert	AV	5.0±0.4	12.4±0.6	21.9±0.3	23.7±0.1	24.1±0.2	17.4	24.5
En-Pt	Cascaded	AV	5.8±0.4	13.8±0.6	23.5±0.4	25.8±0.2	26.3±0.1	19.0	26.4
	AV-Hubert	AV	6.1±0.3	15.5±0.4	26.0±0.3	28.2±0.3	28.6±0.2	20.9	28.8

• Other Cross-Lingual Audio-Visual Speech Tasks

AVMuST-TED can also be used for quantitative verification of cross-lingual multimodal generation tasks, such as Cross-lingual Talking Head Generation.

Download

BaiduDisk 链接: https://pan.baidu.com/s/1Yr6tm56ygEJpCCKe37cAQg?pwd=b9up

License

CC-BY-NC 4.0

Citation

@article{cheng2023mixspeech,
  title={MixSpeech: Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition},
  author={Cheng, Xize and Li, Linjun and Jin, Tao and Huang, Rongjie and Lin, Wang and Wang, Zehan and Liu, Huangdai and Wang, Ye and Yin, Aoxiong and Zhao, Zhou},
  journal={arXiv preprint arXiv:2303.05309},
  year={2023}
}