Give an informative log for whisper on exceptions.

python-api-examples/generate-subtitles.py

@@ -180,6 +180,17 @@ def get_args():
         """,
     )
 
+    parser.add_argument(
+        "--whisper-tail-paddings",
+        default=-1,
+        type=int,
+        help="""Number of tail padding frames.
+        We have removed the 30-second constraint from whisper, so you need to
+        choose the amount of tail padding frames by yourself.
+        Use -1 to use a default value for tail padding.
+        """,
+    )
+
     parser.add_argument(
         "--decoding-method",
         type=str,
@@ -294,6 +305,7 @@ def create_recognizer(args) -> sherpa_onnx.OfflineRecognizer:
             debug=args.debug,
             language=args.whisper_language,
             task=args.whisper_task,
+            tail_paddings=args.whisper_tail_paddings,
         )
     else:
         raise ValueError("Please specify at least one model")

python-api-examples/non_streaming_server.py

@@ -277,6 +277,17 @@ def add_whisper_model_args(parser: argparse.ArgumentParser):
         """,
     )
 
+    parser.add_argument(
+        "--whisper-tail-paddings",
+        default=-1,
+        type=int,
+        help="""Number of tail padding frames.
+        We have removed the 30-second constraint from whisper, so you need to
+        choose the amount of tail padding frames by yourself.
+        Use -1 to use a default value for tail padding.
+        """,
+    )
+
 
 def add_model_args(parser: argparse.ArgumentParser):
     add_transducer_model_args(parser)
@@ -913,6 +924,7 @@ def create_recognizer(args) -> sherpa_onnx.OfflineRecognizer:
             decoding_method=args.decoding_method,
             language=args.whisper_language,
             task=args.whisper_task,
+            tail_paddings=args.whisper_tail_paddings,
         )
     elif args.tdnn_model:
         assert_file_exists(args.tdnn_model)

python-api-examples/offline-decode-files.py

@@ -220,6 +220,17 @@ def get_args():
         """,
     )
 
+    parser.add_argument(
+        "--whisper-tail-paddings",
+        default=-1,
+        type=int,
+        help="""Number of tail padding frames.
+        We have removed the 30-second constraint from whisper, so you need to
+        choose the amount of tail padding frames by yourself.
+        Use -1 to use a default value for tail padding.
+        """,
+    )
+
     parser.add_argument(
         "--decoding-method",
         type=str,
@@ -391,6 +402,7 @@ def main():
             debug=args.debug,
             language=args.whisper_language,
             task=args.whisper_task,
+            tail_paddings=args.whisper_tail_paddings,
         )
     elif args.tdnn_model:
         assert_file_exists(args.tdnn_model)

python-api-examples/two-pass-speech-recognition-from-microphone.py

@@ -195,6 +195,17 @@ def add_second_pass_whisper_model_args(parser: argparse.ArgumentParser):
         """,
     )
 
+    parser.add_argument(
+        "--second-whisper-tail-paddings",
+        default=-1,
+        type=int,
+        help="""Number of tail padding frames.
+        We have removed the 30-second constraint from whisper, so you need to
+        choose the amount of tail padding frames by yourself.
+        Use -1 to use a default value for tail padding.
+        """,
+    )
+
 
 def add_second_pass_non_streaming_model_args(parser: argparse.ArgumentParser):
     add_second_pass_transducer_model_args(parser)
@@ -314,6 +325,7 @@ def create_second_pass_recognizer(args) -> sherpa_onnx.OfflineRecognizer:
             decoding_method="greedy_search",
             language=args.second_whisper_language,
             task=args.second_whisper_task,
+            tail_paddings=args.second_whisper_tail_paddings,
         )
     else:
         raise ValueError("Please specify at least one model for the second pass")

python-api-examples/vad-with-non-streaming-asr.py

@@ -166,6 +166,17 @@ def get_args():
         """,
     )
 
+    parser.add_argument(
+        "--whisper-tail-paddings",
+        default=-1,
+        type=int,
+        help="""Number of tail padding frames.
+        We have removed the 30-second constraint from whisper, so you need to
+        choose the amount of tail padding frames by yourself.
+        Use -1 to use a default value for tail padding.
+        """,
+    )
+
     parser.add_argument(
         "--decoding-method",
         type=str,
@@ -256,6 +267,7 @@ def create_recognizer(args) -> sherpa_onnx.OfflineRecognizer:
             debug=args.debug,
             language=args.whisper_language,
             task=args.whisper_task,
+            tail_paddings=args.whisper_tail_paddings,
         )
     else:
         raise ValueError("Please specify at least one model")

sherpa-onnx/csrc/offline-recognizer-whisper-impl.h

@@ -116,18 +116,12 @@ class OfflineRecognizerWhisperImpl : public OfflineRecognizerImpl {
 
     NormalizeFeatures(f.data(), num_frames, feat_dim);
 
-    // note that 50 is an experience value.
-    // see also ../../scripts/whisper/test.py
-    //
-    // You can replace 50 by other values, say, 100.
+    // note that 1000 is an experience-value.
+    // You can replace 1000 by other values, say, 100.
     //
     // Since we have removed the 30 seconds constraint, we need
     // tail_padding_frames so that whisper is able to detect the eot token.
-    int32_t tail_padding_frames = 50;
-    if (model_->IsMultiLingual()) {
-      // 300 is an experience value. If it throws, please use a larger value.
-      tail_padding_frames = 300;
-    }
+    int32_t tail_padding_frames = 1000;
 
     if (config_.model_config.whisper.tail_paddings > 0) {
       tail_padding_frames = config_.model_config.whisper.tail_paddings;
@@ -140,11 +134,13 @@ class OfflineRecognizerWhisperImpl : public OfflineRecognizerImpl {
 
     Ort::Value mel = Ort::Value::CreateTensor<float>(
         model_->Allocator(), shape.data(), shape.size());
+
     float *p_mel = mel.GetTensorMutableData<float>();
-    std::copy(f.data(), f.data() + actual_frames * feat_dim, p_mel);
+    std::copy(f.data(), f.data() + num_frames * feat_dim, p_mel);
+
+    std::fill_n(p_mel + num_frames * feat_dim,
+                (actual_frames - num_frames) * feat_dim, 0);
 
-    memset(p_mel + f.size(), 0,
-           (actual_frames - num_frames) * feat_dim * sizeof(float));
     mel = Transpose12(model_->Allocator(), &mel);
 
     try {
@@ -156,8 +152,12 @@ class OfflineRecognizerWhisperImpl : public OfflineRecognizerImpl {
       auto r = Convert(results[0], symbol_table_);
       s->SetResult(r);
     } catch (const Ort::Exception &ex) {
-      SHERPA_ONNX_LOGE("\n\nCaught exception:\n\n%s\n\nReturn an empty result",
-                       ex.what());
+      SHERPA_ONNX_LOGE(
+          "\n\nCaught exception:\n\n%s\n\nReturn an empty result. Number of "
+          "input frames: %d, Current tail "
+          "paddings: %d. If you see a lot of such exceptions, please consider "
+          "using a larger --whisper-tail-paddings",
+          ex.what(), num_frames, tail_padding_frames);
       return;
     }
   }

sherpa-onnx/python/sherpa_onnx/offline_recognizer.py

@@ -261,6 +261,7 @@ def from_whisper(
         decoding_method: str = "greedy_search",
         debug: bool = False,
         provider: str = "cpu",
+        tail_paddings: int = -1,
     ):
         """
         Please refer to
@@ -305,6 +306,7 @@ def from_whisper(
                 decoder=decoder,
                 language=language,
                 task=task,
+                tail_paddings=tail_paddings,
             ),
             tokens=tokens,
             num_threads=num_threads,