update

main.py

@@ -112,6 +112,7 @@ def process(
         loader=m_dict["loader"],
         decoding_method=decoding_method,
         num_active_paths=num_active_paths,
+        normalize_samples=m_dict["normalize_samples"],
     )
 
     # transcribe

toolbox/k2_sherpa/nn_models.py

@@ -50,7 +50,19 @@ model_map = {
             "tokens_file_sub_folder": "data/lang_char",
             "loader": "load_sherpa_offline_recognizer",
             "normalize_samples": True,
-        }
+        },
+        {
+            "repo_id": "zrjin/sherpa-onnx-zipformer-multi-zh-hans-2023-9-2",
+            "encoder_model_file": "encoder-epoch-20-avg-1.onnx",
+            "encoder_model_file_sub_folder": ".",
+            "decoder_model_file": "decoder-epoch-20-avg-1.onnx",
+            "decoder_model_file_sub_folder": ".",
+            "joiner_model_file": "joiner-epoch-20-avg-1.onnx",
+            "joiner_model_file_sub_folder": ".",
+            "tokens_file": "tokens.txt",
+            "tokens_file_sub_folder": ".",
+            "loader": "load_sherpa_offline_recognizer_from_transducer",
+        },
     ],
     "English": [
         {
@@ -156,6 +168,30 @@ def load_sherpa_offline_recognizer_from_paraformer(nn_model_file: str,
     return recognizer
 
 
+def load_sherpa_offline_recognizer_from_transducer(encoder_model_file: str,
+                                                   decoder_model_file: str,
+                                                   joiner_model_file: str,
+                                                   tokens_file: str,
+                                                   sample_rate: int = 16000,
+                                                   decoding_method: str = "greedy_search",
+                                                   feature_dim: int = 80,
+                                                   num_threads: int = 2,
+                                                   num_active_paths: int = 2,
+                                                   ):
+    recognizer = sherpa_onnx.OfflineRecognizer.from_transducer(
+        encoder=encoder_model_file,
+        decoder=decoder_model_file,
+        joiner=joiner_model_file,
+        tokens=tokens_file,
+        num_threads=num_threads,
+        sample_rate=sample_rate,
+        feature_dim=feature_dim,
+        decoding_method=decoding_method,
+        max_active_paths=num_active_paths,
+    )
+    return recognizer
+
+
 def load_recognizer(repo_id: str,
                     nn_model_file: str,
                     nn_model_file_sub_folder: str,
@@ -165,6 +201,7 @@ def load_recognizer(repo_id: str,
                     loader: str,
                     decoding_method: str = "greedy_search",
                     num_active_paths: int = 4,
+                    normalize_samples: bool = False,
                     ):
     if not os.path.exists(local_model_dir):
         download_model(
@@ -185,6 +222,7 @@ def load_recognizer(repo_id: str,
             tokens_file=tokens_file,
             decoding_method=decoding_method,
             num_active_paths=num_active_paths,
+            normalize_samples=normalize_samples,
         )
     elif loader == "load_sherpa_offline_recognizer_from_paraformer":
         recognizer = load_sherpa_offline_recognizer_from_paraformer(