test w/ all example

deepspeech/exps/u2/model.py

@@ -264,12 +264,12 @@ class U2Trainer(Trainer):
         config.data.manifest = config.data.test_manifest
         # filter test examples, will cause less examples, but no mismatch with training
         # and can use large batch size , save training time, so filter test egs now.
-        # config.data.min_input_len = 0.0  # second
-        # config.data.max_input_len = float('inf')  # second
-        # config.data.min_output_len = 0.0  # tokens
-        # config.data.max_output_len = float('inf')  # tokens
-        # config.data.min_output_input_ratio = 0.00
-        # config.data.max_output_input_ratio = float('inf')
+        config.data.min_input_len = 0.0  # second
+        config.data.max_input_len = float('inf')  # second
+        config.data.min_output_len = 0.0  # tokens
+        config.data.max_output_len = float('inf')  # tokens
+        config.data.min_output_input_ratio = 0.00
+        config.data.max_output_input_ratio = float('inf')
 
         test_dataset = ManifestDataset.from_config(config)
         # return text ord id

examples/librispeech/s1/README.md

@@ -21,7 +21,6 @@
 | --- | --- | --- | --- | --- | --- | --- | --- |
 | conformer | 47.63 M | conf/conformer.yaml | spec_aug + shift | test-clean-all | attention | 6.35 | 0.057117 |  
 
-
 ## Chunk Conformer
 | Model | Params | Config | Augmentation| Test set | Decode method | Chunk Size & Left Chunks | Loss | WER |  
 | --- | --- | --- | --- | --- | --- | --- | --- | --- |  
@@ -40,3 +39,4 @@
 | Model | Params | Config | Augmentation| Test set | Decode method | Loss | WER |  
 | --- | --- | --- | --- | --- | --- | --- | --- |
 | transformer | 32.52 M | conf/transformer.yaml | spec_aug + shift | test-clean-all | attention | 6.98 | 0.066500 |  
+| transformer | 32.52 M | conf/transformer.yaml | spec_aug + shift | test-clean-all | attention | 7.63 | 0.056832 |  

examples/librispeech/s1/conf/transformer.yaml

@@ -4,7 +4,7 @@ data:
   dev_manifest: data/manifest.dev
   test_manifest: data/manifest.test-clean
   min_input_len: 0.5  # second
-  max_input_len: 20.0 # second
+  max_input_len: 30.0 # second
   min_output_len: 0.0 # tokens
   max_output_len: 400.0 # tokens
   min_output_input_ratio: 0.05

examples/librispeech/s1/run.sh

@@ -5,7 +5,7 @@ source path.sh
 stage=0
 stop_stage=100
 conf_path=conf/transformer.yaml
-avg_num=30
+avg_num=5
 source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
 
 avg_ckpt=avg_${avg_num}

examples/tiny/s1/conf/transformer.yaml

@@ -14,7 +14,7 @@ collator:
   mean_std_filepath: ""
   vocab_filepath: data/vocab.txt 
   unit_type: 'spm'
-  spm_model_prefix: 'data/bpe_unigram_202'
+  spm_model_prefix: 'data/bpe_unigram_200'
   augmentation_config: conf/augmentation.json
   batch_size: 4
   raw_wav: True  # use raw_wav or kaldi feature