Add Export/inference without frontend module

espnet_onnx/asr/model/encoders/encoder.py

@@ -29,7 +29,11 @@ def __init__(
                 self.config.model_path, providers=providers
             )
 
-        self.frontend = Frontend(self.config.frontend, providers, use_quantized)
+        if self.config.frontend.frontend_type is None:
+            self.frontend = None
+        else:
+            self.frontend = Frontend(self.config.frontend, providers, use_quantized)
+
         if self.config.do_normalize:
             if self.config.normalize.type == "gmvn":
                 self.normalize = GlobalMVN(self.config.normalize)
@@ -51,7 +55,11 @@ def __call__(
             speech_lengths: (Batch, )
         """
         # 1. Extract feature
-        feats, feat_length = self.frontend(speech, speech_length)
+        if self.frontend is not None:
+            feats, feat_length = self.frontend(speech, speech_length)
+        else:
+            feats = speech
+            feat_length = speech_length
 
         # 2. normalize with global MVN
         if self.config.do_normalize:

espnet_onnx/export/asr/get_config.py

@@ -90,6 +90,10 @@ def get_frontend_config(asr_frontend_model, frontend=None, **kwargs):
         frontend_config = frontend.get_model_config(**kwargs)
     elif isinstance(asr_frontend_model, DefaultFrontend):
         frontend_config = get_default_frontend(asr_frontend_model)
+    elif asr_frontend_model is None:
+        frontend_config = {
+            "frontend_type": None,
+        }
     else:
         raise ValueError("Currently only s3prl is supported.")
 

tests/integration_tests/test_asr.py

@@ -2,6 +2,7 @@
 
 import librosa
 import pytest
+import numpy as np
 
 from espnet_onnx import Speech2Text as onnxSpeech2Text
 
@@ -21,6 +22,10 @@
     "original/conformer"
 ]
 
+asr_without_frontend = [
+    "original/conformer"
+]
+
 
 @pytest.mark.parametrize("asr_config_names", asr_config_names)
 def test_asr(asr_config_names, load_config, wav_files, model_export):
@@ -78,3 +83,33 @@ def test_asr_custom_dir(asr_config_names, load_config, wav_files, custom_dir_mod
         onnx_output = onnx_model(y)[0]
 
         assert espnet_output[2] == onnx_output[2]
+
+
+@pytest.mark.parametrize("asr_config_names", asr_without_frontend)
+def test_asr_without_frontend(asr_config_names, load_config, custom_dir_model_export):
+    config = load_config(asr_config_names, model_type="integration")
+    config.tag_name = "test/integration/" + config.tag_name
+
+    # build ASR model
+    espnet_model = build_model(config.model_config)
+    espnet_model.asr_model.frontend = None
+
+    # test export
+    export_model(custom_dir_model_export, copy.deepcopy(espnet_model), config)
+    file_paths = check_models(
+        custom_dir_model_export.cache_dir,
+        config.tag_name,
+        config.check_export,
+        False,
+    )
+    check_optimize(config, file_paths)
+    eval_model(espnet_model)
+
+    # parity check with espnet model
+    onnx_model = onnxSpeech2Text(config.tag_name, cache_dir=custom_dir_model_export.cache_dir)
+    for feat_length in [100, 200]:
+        y = np.random.rand(feat_length, 80)
+        espnet_output = espnet_model(y)[0]
+        onnx_output = onnx_model(y)[0]
+
+        assert espnet_output[2] == onnx_output[2]