[to #42322933] fix bugs: audio fs, asr & sv demo services

(cherry picked from commit 1b300b2f3e)

modelscope/pipelines/audio/asr_inference_pipeline.py

@@ -203,13 +203,11 @@ class AutomaticSpeechRecognitionPipeline(Pipeline):
         code_base = self.cmd['code_base']
         self.recog_type = recog_type
         self.audio_format = audio_format
-        self.audio_fs = audio_fs
+        self.audio_fs = None
         checking_audio_fs = None
         self.raw_inputs = None
         if output_dir is not None:
             self.cmd['output_dir'] = output_dir
-        if audio_fs is not None:
-            self.cmd['fs']['audio_fs'] = audio_fs
         self.cmd['param_dict'] = param_dict
 
         if code_base == 'funasr':
@@ -254,6 +252,10 @@ class AutomaticSpeechRecognitionPipeline(Pipeline):
                 self.audio_in, self.audio_format)
             if checking_audio_fs is not None:
                 self.audio_fs = checking_audio_fs
+        if audio_fs is not None:
+            self.cmd['fs']['audio_fs'] = audio_fs
+        else:
+            self.cmd['fs']['audio_fs'] = self.audio_fs
 
         output = self.preprocessor.forward(self.model_cfg, self.recog_type,
                                            self.audio_format, self.audio_in,
@@ -293,7 +295,7 @@ class AutomaticSpeechRecognitionPipeline(Pipeline):
             'seed': 0,
             'ngram_weight': 0.9,
             'nbest': 1,
-            'num_workers': 1,
+            'num_workers': 0,
             'vad_infer_config': None,
             'vad_model_file': None,
             'vad_cmvn_file': None,

modelscope/pipelines/audio/lm_infer_pipeline.py

@@ -156,7 +156,7 @@ class LanguageModelPipeline(Pipeline):
             'dtype': 'float32',
             'ngpu': 1,  # 0: only CPU, ngpu>=1: gpu number if cuda is available
             'seed': 0,
-            'num_workers': 1,
+            'num_workers': 0,
             'log_level': 'ERROR',
             'key_file': None,
             'train_config': lm_model_config,

modelscope/pipelines/audio/punctuation_processing_pipeline.py

@@ -107,7 +107,7 @@ class PunctuationProcessingPipeline(Pipeline):
             'dtype': 'float32',
             'ngpu': 1,  # 0: only CPU, ngpu>=1: gpu number if cuda is available
             'seed': 0,
-            'num_workers': 1,
+            'num_workers': 0,
             'log_level': 'ERROR',
             'key_file': None,
             'train_config': punc_model_config,

modelscope/pipelines/audio/speaker_verification_pipeline.py

@@ -168,8 +168,13 @@ class SpeakerVerificationPipeline(Pipeline):
                 else:
                     audio_scp_1, audio_scp_2 = generate_sv_scp_from_url(
                         audio_in)
-                    data_cmd = [(audio_scp_1, 'speech', 'sound'),
-                                (audio_scp_2, 'ref_speech', 'sound')]
+                    if isinstance(audio_scp_1, bytes) and isinstance(
+                            audio_scp_2, bytes):
+                        data_cmd = [(audio_scp_1, 'speech', 'bytes'),
+                                    (audio_scp_2, 'ref_speech', 'bytes')]
+                    else:
+                        data_cmd = [(audio_scp_1, 'speech', 'sound'),
+                                    (audio_scp_2, 'ref_speech', 'sound')]
             # for raw bytes inputs
             elif isinstance(audio_in[0], bytes):
                 data_cmd = [(audio_in[0], 'speech', 'bytes'),

modelscope/utils/audio/audio_utils.py

@@ -179,7 +179,7 @@ def generate_scp_from_url(url: str, key: str = None):
     if os.path.exists(url) and (url.lower().endswith('.wav')):
         wav_scp_path = url
         return wav_scp_path, raw_inputs
-    # for wav url, download and generate wav.scp
+    # for wav url, download bytes data
     result = urlparse(url)
     if result.scheme is not None and len(result.scheme) > 0:
         storage = HTTPStorage()
@@ -243,17 +243,8 @@ def generate_scp_for_sv(url: str, key: str = None):
     result = urlparse(url)
     if result.scheme is not None and len(result.scheme) > 0:
         storage = HTTPStorage()
-        data = storage.read(url)
-        work_dir = tempfile.TemporaryDirectory().name
-        if not os.path.exists(work_dir):
-            os.makedirs(work_dir)
-        wav_path = os.path.join(work_dir, os.path.basename(url))
-        with open(wav_path, 'wb') as fb:
-            fb.write(data)
-        wav_scp_path = os.path.join(work_dir, 'wav.scp')
-        with open(wav_scp_path, 'w') as ft:
-            scp_content = '\t'.join([wav_name, wav_path]) + '\n'
-            ft.writelines(scp_content)
+        wav_scp_path = storage.read(url)
+        return wav_scp_path
 
     return wav_scp_path
 

requirements/audio/audio_asr.txt

@@ -1,2 +1,2 @@
 easyasr>=0.0.2
-funasr>=0.1.7
+funasr>=0.2.0