Using hamming window for Paraformer frontend.

wenet-e2e · Jun 3, 2024 · aa3eba8 · aa3eba8
1 parent e197305
commit aa3eba8
Show file tree

Hide file tree

Showing 3 changed files with 7 additions and 3 deletions.
diff --git a/wenet/cli/paraformer_model.py b/wenet/cli/paraformer_model.py
@@ -40,7 +40,8 @@ def transcribe(self, audio_file: str, tokens_info: bool = False) -> dict:
                             frame_length=25,
                             frame_shift=10,
                             energy_floor=0.0,
-                            sample_frequency=self.resample_rate)
+                            sample_frequency=self.resample_rate,
+                            window_type="hamming")
         feats = feats.unsqueeze(0)
         feats_lens = torch.tensor([feats.size(1)],
                                   dtype=torch.int64,

diff --git a/wenet/dataset/processor.py b/wenet/dataset/processor.py
@@ -231,7 +231,8 @@ def compute_fbank(sample,
                   num_mel_bins=23,
                   frame_length=25,
                   frame_shift=10,
-                  dither=0.0):
+                  dither=0.0,
+                  window_type="povey"):
     """ Extract fbank
 
         Args:
@@ -253,7 +254,8 @@ def compute_fbank(sample,
                       frame_shift=frame_shift,
                       dither=dither,
                       energy_floor=0.0,
-                      sample_frequency=sample_rate)
+                      sample_frequency=sample_rate,
+                      window_type=window_type)
     sample['feat'] = mat
     return sample
 

diff --git a/wenet/paraformer/convert_paraformer_to_wenet_config_and_ckpt.py b/wenet/paraformer/convert_paraformer_to_wenet_config_and_ckpt.py
@@ -140,6 +140,7 @@ def convert_to_wenet_yaml(configs, wenet_yaml_path: str,
     configs['dataset_conf']['fbank_conf']['frame_shift'] = 10
     configs['dataset_conf']['fbank_conf']['frame_length'] = 25
     configs['dataset_conf']['fbank_conf']['dither'] = 0.1
+    configs['dataset_conf']['fbank_conf']['window_type'] = 'hamming'
     configs['dataset_conf']['spec_sub'] = False
     configs['dataset_conf']['spec_trim'] = False
     configs['dataset_conf']['shuffle'] = True