support asr new models

Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10919277

* support new asr paraformer model

* support asr conformer model

modelscope/pipelines/audio/asr_inference_pipeline.py

@@ -110,6 +110,7 @@ class AutomaticSpeechRecognitionPipeline(Pipeline):
             'sampled_lengths': 'seq2seq/sampled_lengths',
             'lang': 'zh-cn',
             'code_base': inputs['code_base'],
+            'mode': inputs['mode'],
             'fs': {
                 'audio_fs': inputs['audio_fs'],
                 'model_fs': 16000
@@ -233,15 +234,16 @@ class AutomaticSpeechRecognitionPipeline(Pipeline):
     def run_inference(self, cmd):
         asr_result = []
         if self.framework == Frameworks.torch and cmd['code_base'] == 'funasr':
-            from funasr.bin import asr_inference_paraformer_modelscope
+            if cmd['mode'] == 'asr':
+                from funasr.bin import asr_inference_modelscope as asr_inference
+            else:
+                from funasr.bin import asr_inference_paraformer_modelscope as asr_inference
 
-            if hasattr(asr_inference_paraformer_modelscope, 'set_parameters'):
-                asr_inference_paraformer_modelscope.set_parameters(
-                    sample_rate=cmd['fs'])
-                asr_inference_paraformer_modelscope.set_parameters(
-                    language=cmd['lang'])
+            if hasattr(asr_inference, 'set_parameters'):
+                asr_inference.set_parameters(sample_rate=cmd['fs'])
+                asr_inference.set_parameters(language=cmd['lang'])
 
-            asr_result = asr_inference_paraformer_modelscope.asr_inference(
+            asr_result = asr_inference.asr_inference(
                 batch_size=cmd['batch_size'],
                 maxlenratio=cmd['maxlenratio'],
                 minlenratio=cmd['minlenratio'],

modelscope/preprocessors/asr.py

@@ -103,6 +103,12 @@ class WavToScp(Preprocessor):
         else:
             code_base = None
         inputs['code_base'] = code_base
+        # decoding mode
+        if 'mode' in inputs['model_config']:
+            mode = inputs['model_config']['mode']
+        else:
+            mode = None
+        inputs['mode'] = mode
 
         if inputs['model_type'] == Frameworks.torch:
             assert inputs['model_config'].__contains__(
@@ -111,8 +117,6 @@ class WavToScp(Preprocessor):
                 'am_model_config'), 'am_model_config does not exist'
             assert inputs['model_config'].__contains__(
                 'asr_model_config'), 'asr_model_config does not exist'
-            assert inputs['model_config'].__contains__(
-                'asr_model_wav_config'), 'asr_model_wav_config does not exist'
 
             am_model_config: str = os.path.join(
                 inputs['model_workspace'],
@@ -127,9 +131,14 @@ class WavToScp(Preprocessor):
             assert os.path.exists(
                 asr_model_config), 'asr_model_config does not exist'
 
-            asr_model_wav_config: str = os.path.join(
-                inputs['model_workspace'],
-                inputs['model_config']['asr_model_wav_config'])
+            if 'asr_model_wav_config' in inputs['model_config']:
+                asr_model_wav_config: str = os.path.join(
+                    inputs['model_workspace'],
+                    inputs['model_config']['asr_model_wav_config'])
+            else:
+                asr_model_wav_config: str = os.path.join(
+                    inputs['model_workspace'],
+                    inputs['model_config']['asr_model_config'])
             assert os.path.exists(
                 asr_model_wav_config), 'asr_model_wav_config does not exist'
 

requirements/audio.txt

@@ -1,6 +1,6 @@
 easyasr>=0.0.2
 espnet==202204
-funasr>=0.1.0
+funasr>=0.1.3
 h5py
 inflect
 keras

tests/pipelines/test_automatic_speech_recognition.py

@@ -217,6 +217,41 @@ class AutomaticSpeechRecognitionTest(unittest.TestCase,
             'damo/speech_UniASR_asr_2pass-id-16k-common-vocab1067-tensorflow1-offline',
             'wav_path': 'data/test/audios/asr_example_id.wav'
         },
+        {
+            'model_id':
+            'damo/speech_conformer_asr_nat-zh-cn-16k-aishell1-vocab4234-pytorch',
+            'wav_path': 'data/test/audios/asr_example_id.wav'
+        },
+        {
+            'model_id':
+            'damo/speech_conformer_asr_nat-zh-cn-16k-aishell2-vocab5212-pytorch',
+            'wav_path': 'data/test/audios/asr_example_id.wav'
+        },
+        {
+            'model_id':
+            'damo/speech_paraformer_asr_nat-zh-cn-16k-aishell1-vocab4234-pytorch',
+            'wav_path': 'data/test/audios/asr_example_id.wav'
+        },
+        {
+            'model_id':
+            'damo/speech_paraformer_asr_nat-zh-cn-16k-aishell2-vocab5212-pytorch',
+            'wav_path': 'data/test/audios/asr_example_id.wav'
+        },
+        {
+            'model_id':
+            'damo/speech_paraformerbert_asr_nat-zh-cn-16k-aishell1-vocab4234-pytorch',
+            'wav_path': 'data/test/audios/asr_example_id.wav'
+        },
+        {
+            'model_id':
+            'damo/speech_paraformerbert_asr_nat-zh-cn-16k-aishell2-vocab5212-pytorch',
+            'wav_path': 'data/test/audios/asr_example_id.wav'
+        },
+        {
+            'model_id':
+            'damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch',
+            'wav_path': 'data/test/audios/asr_example_id.wav'
+        },
     ]
 
     def setUp(self) -> None: