Shengqiang Li

Issue about performance

8

I have finished training and decoding in AISHELL-1 dataset and got cer=12.4% in test set，and i found that my model.json which uses the default config is different from the one...

About the experimental settings in the paper "HIFI-CODEC: GROUP-RESIDUAL VECTOR QUANTIZATION FOR HIGH FIDELITY AUDIO CODEC"

1

I want to reproduce the results in this paper "HIFI-CODEC: GROUP-RESIDUAL VECTOR QUANTIZATION FOR HIGH FIDELITY AUDIO CODEC". However, the description is so confused. The paper just said that the...

1. 问题：在训练阶段，LLM输入序列中有speaker embedding。在指令微调模型推理时LLM输入序列没有speaker embedding。因此对于指令微调模型而言，训练和推理存在mismatch，效果不理想。 2. 解决方案：LLM增加instruct_finetuning的参数，缺省值为False。当该参数为True时，训练阶段也会去掉LLM输入序列中的speaker embedding，从而保证指令微调模型可以正常训练和推理。