基于数据驱动缺失特征检测与重建的声纹识别

摘要

Pointing at improving the noise robustness of speaker recognition systems, a front-end processing method for voiceprint recognition based on data training missing feature detection and reconstruction is proposed. Taking full advantage of the a priori information obtained from massive training data, this method can accurately calculate subband SNR, detect, mark and reconstruct the subband feature parameters which are serious polluted by various noise, in this way, noise robust feature parameter is obtained. Experiments show that in low SNR environment, data training method has achieved a high recognition rate, and system performance also improved in none stationary environment.%声纹识别系统的识别性能会随着环境噪声的增强而急剧降低，为了使系统具备一定的噪声鲁棒性，提出了一种基于数据驱动缺失特征检测与重建的声纹识别前端处理方法。充分利用大量数据训练得到的信息估计子带信噪比，检测、标记和重建被噪声污染严重的子带特征，从而得到噪声鲁棒性特征参数。实验表明，该方法在低信噪比环境下取得了较高的识别率提升，在非平稳噪声下系统性能也有着较好的改善。