Use a larger padding value for multilingual whisper models

k2-fsa · Dec 7, 2023 · 191407a · 191407a
1 parent 81f502a
commit 191407a
Show file tree

Hide file tree

Showing 2 changed files with 3 additions and 3 deletions.
diff --git a/scripts/whisper/test.py b/scripts/whisper/test.py
@@ -259,7 +259,7 @@ def compute_features(filename: str) -> torch.Tensor:
     # You can use another value instead of 50.
     mel = torch.nn.functional.pad(mel, (0, 0, 0, 50), "constant", 0)
     # Note that if it throws for a multilingual model,
-    # please use a larger value, say 200
+    # please use a larger value, say 300
 
     target = 3000
     if mel.shape[0] > target:

diff --git a/sherpa-onnx/csrc/offline-recognizer-whisper-impl.h b/sherpa-onnx/csrc/offline-recognizer-whisper-impl.h
@@ -124,8 +124,8 @@ class OfflineRecognizerWhisperImpl : public OfflineRecognizerImpl {
     // tail_padding_frames so that whisper is able to detect the eot token.
     int32_t tail_padding_frames = 50;
     if (model_->IsMultiLingual()) {
-      // 200 is an experience value. If it throws, please use a larger value.
-      tail_padding_frames = 200;
+      // 300 is an experience value. If it throws, please use a larger value.
+      tail_padding_frames = 300;
     }
 
     int32_t actual_frames =