What it is

We extract code from SLAM-LLM and FireredASR to support the fine-tuning of the FireredASR-LLM model.

Finetuning

./scripts/finetune_deepspeed.sh

Inference

./scripts/decode_deepspeed.sh

Features

Large-scale industrial dataset: adopt an iterative dataset and dynamic batching strategy.

On-the-fly data augmentation: support spectral augmentaton, reverberation and noise adding.

Training on NPU: multi-machine multi-NPU training using deepspeed. If GPU training is required, simply replace all .npu with .cuda in the code.