About Modality Corpus

high-resolution, high-framerate stereoscopic video streams and audio signals

Corpus Features

01

35 speakers

including 17 natives and 18 non-natives

02

2.1 TB

of high quality audio-visual material

03

Full HD / 100 FPS

video capture

04

8 PCM audio streams

gathered from a microphone array

05

Different recording conditions

includes recordings in clean and noisy conditions

06

Commands/sentences

Includes separated commands and continuous sentences

07

Labeled material

Corpus contains hand-made label files as ground truth for AVSR algorithms

About files formats

The audio files use the Waveform Audio File Format (.wav), and contain a single PCM audio stream sampled at 44.1 kSa/s with 16-bit depth. The video files utilize the Matroska Multimedia Container Format (.mkv) in which a video stream in 1080p resolution, captured at 100 fps was placed after being compressed with h.264 codec (using High 4:4:4 profile). The ‘.lab’ files are text files containing the information on word positions in audio files, and follow the HTK label format. Each line of a ‘.lab’ file contains the actual label preceded by start and end times (in 100 ns units) e.g. : 1239620000 1244790000 FIVE which denotes the word “five”, occurring between the 123.962 s and 124.479 s of audio.

If you like to read

The MODALITY corpus consists of over 30 hours of multimodal recordings. The database contains high-resolution, high-framerate stereoscopic video streams and audio signals obtained from a microphone array and a laptop microphone. The corpus can be employed to develop an AVSR system, as every utterance was labelled. Recordings in noisy conditions can be used to test the robustness of speech recognition systems.

The language material was based on a remote control scenario and it includes 231 words -numbers, names of months and days, a set of verbs and nouns related to a computer device control. They were read by speakers as separated words and sequences resulting in a set of 12 recording sessions per speaker. Half of the sessions were recorded in quiet conditions, the other half contained three kinds of intrusive signals (traffic, babble and factory noise).

The corpus includes recordings of 42 speakers (33 male, 9 female). The participants include 20 students and staff of Multimedia Systems Department of the Gdańsk University of Technology, 5 students of the Institute of English and American Studies of the University of Gdańsk, and 17 native English speakers.

Due to the size of the corpus (approx. 2.5 TB of data), every speaker’s recording was placed in a separate zip file of the size approx. 4-7 GB each. Multiple, chosen recordings can be grouped into a single archive to ease the process of downloading the corpus.

The recordings were organized according to the speakers’ language skills. The group A (17 speakers) consists of native-speakers. Non-native speakers recordings (Polish nationals) were placed in the Group B (25 speakers).

The MODALITY is described in a more detailed way in this paper:
Piotr Bratoszewski, Andrzej Czyżewski, Józef Kotus, Paweł Spaleniak, Marcin
Title: An audio-visual corpus for multimodal automatic speech recognition (currently under review for Journal of Intelligent Information Systems )