Our team at Microsoft recently organized the
Deep Noise Suppression Challenge at INTERSPEECH 2020 to help stimulate research in this important area of research. The response from that challenge exceeded our expectations with 19 teams submitting 28 models for noise suppression. The dataset and test
framework continue to be cloned and forked at a good rate since the challenge, indicating it has good utility for the research community.
Another important area for speech enhancement is acoustic echo cancellation, which is still a top issue in audio communication and conferencing systems. We are submitting a proposal to organize an Acoustic Echo Cancellation Challenge at
ICASSP 2021 to stimulate research in this area. The challenge will have an open source dataset for acoustic echo cancellation and detection, captured using a large-scale crowdsourcing effort. We will provide a dataset of 10,000 different real environments,
audio devices, and human speakers in the following scenarios:
-
Far-end single talk, no echo path change
-
Far-end single talk, echo path change
-
Near-end single talk, no echo path change
-
Double talk, no echo path change
-
Double talk, echo path change
-
Sweep signal for RT60 estimation
Each scenario includes the speaker, microphone, and loopback signal. Besides the real data, we will also provide synthetic data which is useful for training.
We propose two tracks:
-
Real-time acoustic echo detection (AED) classifier.
-
Real-time acoustic echo cancellation (AEC).
In both the tracks:
-
The AED/AEC must run in less than T/2 (in ms) to process a frame of size T (in ms) on an Intel Core i5 quad-core machine clocked at 2.4 GHz or equivalent processors. Frame length T must be less than or equal to 40ms (0<T<=40ms).
-
The AED/AEC method may have a maximum of 40ms look ahead. To infer the current frame T (in ms), the algorithm can access any number of past frames but only 40ms of future frames (T+40ms).
-
We will provide a base model implementation.
-
We will provide an objective and subjective evaluation framework.
-
The AED/AEC can be a deep model, a traditional signal processing algorithm, or a mix of the two. There are no restrictions on the AED/AEC aside from the run time and receptive field described above.
Participants can enter either one or both tracks. Prizes will be awarded to the top three participants in each track.
For the proposal, ICASSP requires us to provide them the list of potential participants with their potential title and author(s) details. If you are interested to participate in this challenge, please send an email to
ross.cutler@xxxxxxxxxxxxx the following details by August 9th.
-
A potential title of your paper. This can be updated later.
-
A potential list of authors. This can be updated later.
-
A webpage/URL of your personal webpage, LinkedIn page, or your research lab page.
We look forward to hearing back from you and we hope to see you participate in the challenge.
Regards,
Acoustic Echo Cancellation Challenge Organizers