Interspeech 2021 Deep Noise Suppression Challenge

深度噪声抑制 (DNS) 挑战旨在促进噪声抑制领域的创新，以实现卓越的感知语音质量。这个挑战有两条两条轨道： 轨道 1：宽带场景的实时去噪轨道 噪声抑制器必须花费少于步幅时间 Ts（以毫秒为单位）来处理英特尔酷睿 i5 上大小为 T（以毫秒为单位）的帧主频为 2.4 GHz 的四核机器或同等处理器。例如，对于帧之间 50% 的重叠，Ts = T/2。允许的总算法延迟包括帧大小 T、步幅时间 Ts 和任何前瞻必须小于或等于 40 毫秒。例如，对于接收 20 毫秒音频块的实时系统，如果您使用 20 毫秒的帧长度和 10 毫秒的步幅，从而导致算法延迟为 30 毫秒，那么您就满足了延迟要求。如果您使用大小为 32 毫秒且步长为 16 毫秒的帧，导致算法延迟为 48 毫秒，那么您的方法不满足延迟要求，因为总算法延迟超过 40 毫秒。如果您的帧大小加上步幅 T1=T+Ts 小于 40 毫秒，那么您最多可以使用 (40-T1) 毫秒的未来信息。轨道 2：全频段场景的实时去噪轨道 满足轨道 1 的要求，但频率为 48 kHz。

The Deep Noise Suppression (DNS) Challenge aims to promote innovation in the field of noise suppression to achieve superior perceptual speech quality. This challenge includes two tracks:

Track 1: Real-Time Denoising for Wideband Scenarios
Noise suppressors must complete processing of frames with a duration of T (in milliseconds) within a time shorter than the stride time Ts (in milliseconds) on a quad-core Intel Core i5 processor running at 2.4 GHz or an equivalent processor. For example, with 50% overlap between consecutive frames, Ts = T/2. The total allowable algorithm latency, which includes the frame size T, stride time Ts, and any lookahead, must be less than or equal to 40 milliseconds. For instance, for a real-time system receiving 20-millisecond audio chunks, if a frame length of 20 ms and a stride of 10 ms are used, resulting in an algorithm latency of 30 ms, the latency requirement is satisfied. If frames of 32 ms with a stride of 16 ms are used instead, resulting in an algorithm latency of 48 ms, the method fails to meet the latency requirement, as the total algorithm latency exceeds 40 ms. If T1 = T + Ts (sum of frame size and stride time) is less than 40 ms, a maximum of (40 - T1) milliseconds of future audio information may be utilized.

Track 2: Real-Time Denoising for Full-band Scenarios
This track satisfies all requirements specified in Track 1, but operates at a sampling frequency of 48 kHz.

搜集汇总

背景与挑战

背景概述

Interspeech 2021 Deep Noise Suppression Challenge是一个旨在促进噪声抑制领域创新的数据集，包含宽带和全频段场景的实时去噪两条轨道，要求算法延迟不超过40毫秒。该数据集由微软于2021年发布，相关论文可在arXiv上找到。