save

Author: 蒄骰

examples/contextual_asr/README.md

@@ -28,15 +28,21 @@ words, with the remainder classified as rare words. The biasing list generated f
 
 
 ## Decoding with checkpoints
+LLM-based Contextual ASR Inference script, with different biaisng sizes and test sets.
 ```
-bash decode_wavlm_libri960_ft_char_hotwords_filter_N100_F3.sh
+bash decode_wavlm_libri960_ft_char_hotwords.sh
 ```
 
 ## Training the model
+LLM-based ASR Training script: using CTC fine-tuned Wavlm as encoder and “Transcribe speech to text.” as prompt.
 ```
-bash finetune_wavlm_libri960_ft_char_hotwords.sh
 bash finetune_wavlm_libri960_ft_char.sh
 ```
+LLM-based Contextual ASR Training script: using CTC fine-tuned Wavlm as encoder and "Transcribe speech to text. Some hotwords might help. The hotwords are {}.” as prompt.
+```
+bash finetune_wavlm_libri960_ft_char_hotwords.sh
+```
+
 
 ##  Citation
 You can refer to the paper for more results. 

examples/contextual_asr/docs/model.pdf

@@ -17,8 +17,8 @@ output_dir=/nfs/yangguanrou.ygr/experiments_librispeech/vicuna-7b-v1.5-WavLM-Lar
 ckpt_path=$output_dir/asr_epoch_3_step_25780
 
 first=1
-for N in 100; do
-        for ref_split in test_clean ; do #test_other
+for N in 100 500 1000 2000; do
+        for ref_split in test_clean test_other; do
                 split=librispeech_${ref_split}
                 val_data_path=/nfs/maziyang.mzy/data/librispeech/${split}.jsonl
                 decode_log=$ckpt_path/decode_${split}_beam4_filter_N${N}_first${first}_debug