Voice Bank + DEMAND

该数据集包含了用于语音增强的噪声-清晰音频对，其中训练集由来自28位发言人的11,572对音频组成，测试集则由2位未见发言人的824对音频组成。音频样本在训练时与10种类型的噪声以四个不同的信噪比水平混合，而测试语音则是使用5种未见噪声类型，在不同信噪比水平下创建的。所有音频均以16千赫兹重新采样，并分割成3秒的片段。规模方面，训练集包含11,572对音频，测试集包含824对音频。该数据集的任务是语音增强。

This dataset provides noise-clean audio pairs for speech enhancement. The training set comprises 11,572 audio pairs from 28 distinct speakers, whereas the test set contains 824 audio pairs from 2 unseen speakers. During training, clean audio samples were mixed with 10 types of noise at four different signal-to-noise ratio (SNR) levels; the test noisy speech was generated using 5 unseen noise types across varying SNR levels. All audio recordings were resampled to 16 kHz and segmented into 3-second clips. In terms of dataset scale, the training set includes 11,572 audio pairs and the test set contains 824 audio pairs. The core task of this dataset is speech enhancement.

搜集汇总

背景与挑战

背景概述

Voice Bank + DEMAND数据集用于单通道语音增强研究，DB-AIAT模型在该数据集上实现了3.31 PESQ、95.6% STOI和10.79dB SSNR的性能，模型轻量（2.81M）。