Two-channel mixtures of speech and real-world background noise

This task aims at evaluating source separation and denoising techniques in the context of speech enhancement by merging two datasets: the SiSEC 2010 noisy speech dataset and the CHiME corpus. Both datasets consist of two-channel mixtures of one speech source and real-world background noise, so that algorithms applicable to one dataset are applicable to the other without additional effort. The source separation results obtained over the latter dataset will be analyzed in line of the speech recognition results obtained over that dataset as part of the CHiME Challenge.

Description of the dataset

We consider two-channel mixtures of one speech source and real-world background noise sampled at 16 kHz. The dataset consists of two parts:

Public environments

These data are part of the SiSEC 2010 noisy speech dataset. Background noise signals were recorded via a pair of omnidirectional microphones spaced by 8.6 cm in six different public environments:

• Su1: subway car moving
• Su2: subway car standing at station
• Ca1: cafeteria 1
• Ca2: cafeteria 2
• Sq1: square 1
• Sq2: square 2

and in two different positions within each environment:

• Ce: center (except in Su)
• Co: corner

Several recordings identified by letters (A or B) were made in each case. Mixtures were then generated by adding a speech signal to the background noise signal. For the reverberant environments Su and Ca, the speech signals were recorded in an office room using the same microphone pair. For the outdoor environment Sq, the speech signals were mixed anechoically. The distance between the sound source and the array centroid was 1.0 m for female speech and 0.8 m for male speech. The direction of arrival (DOA) of the speech source was different in each mixture and the signal-to-noise ratio (SNR) was drawn randomly between -17 and +12 dB.

Domestic environment

These data are part of the CHiME corpus. Background noise signals were recorded via a pair of in-ear microphones placed on a mannikin in a single domestic environment:

• Li: living room

at six different SNRs:

• -6 dB
• -3 dB
• 0 dB
• 3 dB
• 6 dB
• 9 dB

Several recordings were made for a duration of 5 min each. Mixtures were then generated by adding a number of speech sentences to the background noise signal. The speech signals were generated by convolving clean speech sentences with binaural room impulse responses (BRIR) measured at a position 2 m directly in front of the mannikin. Hence the DOA of the speech source is known and fixed. Among this wealth of data, we selected four successive sentences identified by letters (A to D) for each SNR. The silence between successive sentences was removed. Entrants can either process each sentence separately, exploit the neighboring sentences or the whole 5 min context. For more details about this dataset, see the following paper.

Test data

Download the test set (13 MB)

The data consist of 44 stereo WAV audio files that can be imported in Matlab using the wavread command. These files are named test_<env>_<cond>_<take>_mix.wav, where <env> is the noise environment, <cond> the recording condition (center/corner or SNR) and <take> a letter from A to D.

Entrants wishing to exploit the context of each sentence in the domestic environment database can also download the corresponding 5 min recordings (87 MB)

The data consist of 6 stereo WAV audio files that can be imported in Matlab using the wavread command. The text file embedding.txt describes the position of the selected sentences within these recordings. Each line provides the following five pieces of information:

• sentence filename (SiSEC)
• 5 min recording filename
• start position in samples within the 5 min recording
• duration in samples
• sentence filename (ChiME)

Development data

Download the development set (24 MB)

The data consists of 136 WAV audio files that can be imported in Matlab using the wavread command and 10 text files. These files are named as follows:

• dev_<env>_<cond>_<take>_src.wav: single-channel speech signal
• dev_<env>_<cond>_<take>_sim.wav: two-channel spatial image of the speech source
• dev_<env>_<cond>_<take>_noi.wav: two-channel spatial image of the background noise
• dev_<env>_<cond>_<take>_mix.wav: two-channel mixture signal
• dev_<env>_<cond>_<take>_DOA.txt: DOA of the speech source (see the SiSEC 2010 wiki for the convention adopted to measure DOA)

Since the source DOAs were measured geometrically in the Su and Ca environments, they might contain a measurement error up to a few degrees; on the contrary, there is no such error in the Sq environment.

Entrants wishing to exploit the context of each sentence in the domestic environment database can also download the corresponding 5 min recordings (86 MB) (same nomenclature as above).

Training data

The CHiME database also includes training data for each speaker and for the background noise.

Tasks and reference software

We propose the following 3 tasks:

1. speaker DOA estimation: estimate the DOA of the speech source (public environments only)
2. speech signal estimation: estimate the single-channel dereverberated speech signal
3. speech and noise spatial image estimation: decompose the mixture signal into two two-channel signals corresponding to the speech source and the background noise

Participants are welcome to use some of the Matlab reference software below to build their own algorithms:

• stft_multi.m: multichannel STFT
• istft_multi.m: multichannel inverse STFT
• example_denoising.m: TDOA estimation by GCC-PHATmax, ML target and noise variance estimation under a diffuse noise model, and multichannel Wiener filtering

Due to the specific construction of the dataset, at least four strategies may be employed to process the domestic environment mixtures:

1. process each mixture (= 1 isolated sentence) alone
2. process all mixtures with the same SNR (= 4 successive sentences without silence) together
3. process the whole 5 min recording without knowledge of the sentence positions
4. process the whole 5 min recording using knowledge of the sentence positions

In any case, it is expected that the submitted signals correspond to the test mixtures (= isolated sentences).

Submission

Each participant is asked to submit the results of his/her algorithm for task 2 and/or 3 over all or part of the mixtures in the development dataset and the test dataset. The results for task 1 may also be submitted if possible.

Each participant should make his results available online in the form of a tarball with the following file naming convention:

• test_<env>_<cond>_<take>_src.wav: single-channel speech signal
• test_<env>_<cond>_<take>_sim.wav: two-channel spatial image of the speech source
• test_<env>_<cond>_<take>_noi.wav: two-channel spatial image of the background noise
• test_<env>_<cond>_<take>_DOA.txt: DOA of the speech source

For the domestic environment dataset, the CHiME file naming convention is also acceptable.

Each participant should then send an email to "shoko (at) lab.ntt.co.jp", "nesta (a) fbk.eu" and "emmanuel.vincent (at) inria.fr" providing:

• contact information (name, affiliation)
• basic information about his/her algorithm, including the employed processing strategy among the four strategies outlined above, its average running time (in seconds per test excerpt and per GHz of CPU) and a bibliographical reference if possible
• the URL of the tarball

The submitted audio files will be made available on a website under the terms of the Licensing section below.

Evaluation criteria

We propose to use the same evaluation criteria as in SiSEC 2010, except that the order of the estimated sources must be recovered.

The estimated speaker DOAs in task 1 will be evaluated in terms of absolute difference with the true DOAs.

The estimated speech signals in task 2 will be evaluated via the energy ratio criteria defined in the BSS_EVAL toolbox allowing arbitrary filtering between the estimated source and the true source and via the perceptually-motivated criteria in the PEASS toolkit.

The estimated speech and noise spatial image signals in task 3 will be evaluated via the energy ratio criteria introduced for the Stereo Audio Source Separation Evaluation Campaign and via the perceptually-motivated criteria in the PEASS toolkit.

Performance will be compared to that of ideal binary masking as a benchmark (i.e. binary masks providing maximum SDR), computed over a STFT or a cochleagram.

The above performance criteria and benchmarks are respectively implemented in

• bss_eval_source_denoising.m
• bss_eval_images_nosort.m
• PEASS
• sep_ibm.m
• Cochleagram Toolbox

An example use is given in example_denoising.m.

Licensing issues

All files are distributed under the terms of the Creative Commons Attribution-Noncommercial-ShareAlike 3.0 license. The files to be submitted by participants will be made available on a website under the terms of the same license.

Public environment data were authored by Ngoc Q. K. Duong and Nobutaka Ito.

Domestic environment data were authored by H. Christensen, Jon Barker, Ning Ma and Phil Green in the framework of the CHiME Project funded by the EPSRC. If you use these data in any published research please cite:

H. Christensen, J. Barker, N. Ma and P. Green. The CHiME corpus: a resource and a challenge for Computational Hearing in Multisource Environments. In Proc. Interspeech'10, 2010. (pdf).

Potential Participants

• Shoko Araki (araki.shoko (a) lab_ntt_co_jp)
• Dorothea Kolossa (dorothea_kolossa (a) ruhr-uni-bochum_de)
• Alexey Ozerov (alexey_ozerov (a) inria_fr)
• Francesco Nesta (nesta (a) fbk_eu)
• Armin Sehr (sehr (a) nt_e-technik_uni-erlangen_de)
• Ngoc Duong
• Jani Even
• Hiroshi Saruwatari
• Dang Hai Tran Vu
• Hiroshi Sawada

Task proposed by the Audio Committee