Support selecting decoding method.

app.py

@@ -39,7 +39,13 @@ def convert_to_wav(in_filename: str) -> str:
     return out_filename
 
 
-def process(in_filename: str, language: str, repo_id: str) -> str:
+def process(
+    in_filename: str,
+    language: str,
+    repo_id: str,
+    decoding_method: str,
+    num_active_paths: int,
+) -> str:
     print("in_filename", in_filename)
     print("language", language)
     print("repo_id", repo_id)
@@ -65,7 +71,11 @@ def process(in_filename: str, language: str, repo_id: str) -> str:
         )
     wave = wave[0]  # use only the first channel.
 
-    hyp = get_pretrained_model(repo_id).decode_waves([wave])[0]
+    hyp = get_pretrained_model(repo_id).decode_waves(
+        [wave],
+        decoding_method=decoding_method,
+        num_active_paths=num_active_paths,
+    )[0]
 
     date_time = now.strftime("%Y-%m-%d %H:%M:%S.%f")
     end = time.time()
@@ -107,12 +117,12 @@ demo = gr.Blocks()
 
 with demo:
     gr.Markdown(title)
-    gr.Markdown(description)
     language_choices = list(language_to_models.keys())
 
     language_radio = gr.Radio(
         label="Language",
         choices=language_choices,
+        value=language_choices[0],
     )
     model_dropdown = gr.Dropdown(choices=[], label="Select a model")
     language_radio.change(
@@ -121,6 +131,19 @@ with demo:
         outputs=model_dropdown,
     )
 
+    decoding_method_radio = gr.Radio(
+        label="Decoding method",
+        choices=["greedy_search", "modified_beam_search"],
+        value="greedy_search",
+    )
+
+    num_active_paths_slider = gr.Slider(
+        minimum=1,
+        value=4,
+        step=1,
+        label="Number of active paths for modified_beam_search",
+    )
+
     with gr.Tabs():
         with gr.TabItem("Upload from disk"):
             uploaded_file = gr.inputs.Audio(
@@ -149,14 +172,27 @@ with demo:
 
         upload_button.click(
             process,
-            inputs=[uploaded_file, language_radio, model_dropdown],
+            inputs=[
+                uploaded_file,
+                language_radio,
+                model_dropdown,
+                decoding_method_radio,
+                num_active_paths_slider,
+            ],
             outputs=uploaded_output,
         )
         record_button.click(
             process,
-            inputs=[microphone, language_radio, model_dropdown],
+            inputs=[
+                microphone,
+                language_radio,
+                model_dropdown,
+                decoding_method_radio,
+                num_active_paths_slider,
+            ],
             outputs=recorded_output,
         )
+    gr.Markdown(description)
 
 if __name__ == "__main__":
     demo.launch()

offline_asr.py

@@ -223,14 +223,6 @@ class OfflineAsr(object):
           token_filename:
             Path to tokens.txt. If it is None, you have to provide
             `bpe_model_filename`.
-          decoding_method:
-            The decoding method to use. Currently, only greedy_search and
-            modified_beam_search are implemented.
-          num_active_paths:
-            Used only when decoding_method is modified_beam_search.
-            It specifies number of active paths for each utterance. Due to
-            merging paths with identical token sequences, the actual number
-            may be less than "num_active_paths".
           sample_rate:
             Expected sample rate of the feature extractor.
           device:
@@ -254,24 +246,6 @@ class OfflineAsr(object):
             device=device,
         )
 
-        assert decoding_method in (
-            "greedy_search",
-            "modified_beam_search",
-        ), decoding_method
-        if decoding_method == "greedy_search":
-            nn_and_decoding_func = run_model_and_do_greedy_search
-        elif decoding_method == "modified_beam_search":
-            nn_and_decoding_func = functools.partial(
-                run_model_and_do_modified_beam_search,
-                num_active_paths=num_active_paths,
-            )
-        else:
-            raise ValueError(
-                f"Unsupported decoding_method: {decoding_method} "
-                "Please use greedy_search or modified_beam_search"
-            )
-
-        self.nn_and_decoding_func = nn_and_decoding_func
         self.device = device
 
     def _build_feature_extractor(
@@ -300,7 +274,12 @@ class OfflineAsr(object):
 
         return fbank
 
-    def decode_waves(self, waves: List[torch.Tensor]) -> List[List[str]]:
+    def decode_waves(
+        self,
+        waves: List[torch.Tensor],
+        decoding_method: str,
+        num_active_paths: int,
+    ) -> List[List[str]]:
         """
         Args:
           waves:
@@ -314,14 +293,40 @@ class OfflineAsr(object):
               then the given waves have to contain samples in this range.
 
               All models trained in icefall use the normalized range [-1, 1].
+          decoding_method:
+            The decoding method to use. Currently, only greedy_search and
+            modified_beam_search are implemented.
+          num_active_paths:
+            Used only when decoding_method is modified_beam_search.
+            It specifies number of active paths for each utterance. Due to
+            merging paths with identical token sequences, the actual number
+            may be less than "num_active_paths".
         Returns:
           Return a list of decoded results. `ans[i]` contains the decoded
           results for `wavs[i]`.
         """
+        assert decoding_method in (
+            "greedy_search",
+            "modified_beam_search",
+        ), decoding_method
+
+        if decoding_method == "greedy_search":
+            nn_and_decoding_func = run_model_and_do_greedy_search
+        elif decoding_method == "modified_beam_search":
+            nn_and_decoding_func = functools.partial(
+                run_model_and_do_modified_beam_search,
+                num_active_paths=num_active_paths,
+            )
+        else:
+            raise ValueError(
+                f"Unsupported decoding_method: {decoding_method} "
+                "Please use greedy_search or modified_beam_search"
+            )
+
         waves = [w.to(self.device) for w in waves]
         features = self.feature_extractor(waves)
 
-        tokens = self.nn_and_decoding_func(self.model, features)
+        tokens = nn_and_decoding_func(self.model, features)
 
         if hasattr(self, "sp"):
             results = self.sp.decode(tokens)