search

YuanGongND / ssast

Code for the AAAI 2022 paper "SSAST: Self-Supervised Audio Spectrogram Transformer".
BSD 3-Clause "New" or "Revised" License
362 stars 58 forks source link

fine-tuning sample rate #10

Closed skirdey closed 2 years ago

skirdey commented 2 years ago

When using pre-trained models for fine-tuning, shall the fine-tuning training set have a specific sample rate, like 16khz?

YuanGongND commented 2 years ago

The model is pretrained with only 16kHz data (both AudioSet and Librispeech we use to train the model are re-sampled to 16kHz), so my guess is in the fine-tuning stage, the sampling rate should be consistent. Otherwise you can pretrain the model using a different sampling rate, the pretraining is not that expensive (a few days on 4X1080 GPUs).

-Yuan