[binding] add chunk size interface and use non-streaming decoding by …

…default since it's fast and accurate (#1970)

runtime/binding/python/README.md

@@ -91,7 +91,7 @@ with wave.open(test_wav, 'rb') as fin:
     assert fin.getnchannels() == 1
     wav = fin.readframes(fin.getnframes())
 
-decoder = wenet.Decoder(lang='chs')
+decoder = wenet.Decoder(lang='chs', streaming=True)
 # We suppose the wav is 16k, 16bits, and decode every 0.5 seconds
 interval = int(0.5 * 16000) * 2
 for i in range(0, len(wav), interval):

runtime/binding/python/cpp/binding.cc

@@ -18,7 +18,6 @@
 
 namespace py = pybind11;
 
-
 PYBIND11_MODULE(_wenet, m) {
   m.doc() = "wenet pybind11 plugin";  // optional module docstring
   m.def("wenet_init", &wenet_init, py::return_value_policy::reference,
@@ -36,4 +35,6 @@ PYBIND11_MODULE(_wenet, m) {
   m.def("wenet_set_language", &wenet_set_language, "set language");
   m.def("wenet_set_continuous_decoding", &wenet_set_continuous_decoding,
         "enable continuous decoding or not");
+  m.def("wenet_set_chunk_size", &wenet_set_chunk_size,
+        "set decoding chunk size");
 }

runtime/binding/python/wenetruntime/decoder.py

@@ -34,7 +34,8 @@ def __init__(self,
                  enable_timestamp: bool = False,
                  context: Optional[List[str]] = None,
                  context_score: float = 3.0,
-                 continuous_decoding: bool = False):
+                 continuous_decoding: bool = False,
+                 streaming: bool = False):
         """ Init WeNet decoder
         Args:
             lang: language type of the model
@@ -44,6 +45,7 @@ def __init__(self,
             context: context words
             context_score: bonus score when the context is matched
             continuous_decoding: enable countinous decoding or not
+            streaming: streaming mode
         """
         if model_dir is None:
             model_dir = Hub.get_model_by_lang(lang)
@@ -57,6 +59,8 @@ def __init__(self,
             self.add_context(context)
             self.set_context_score(context_score)
         self.set_continuous_decoding(continuous_decoding)
+        chunk_size = 16 if streaming else -1
+        self.set_chunk_size(chunk_size)
 
     def __del__(self):
         _wenet.wenet_free(self.d)
@@ -90,6 +94,9 @@ def set_continuous_decoding(self, continuous_decoding: bool):
         flag = 1 if continuous_decoding else 0
         _wenet.wenet_set_continuous_decoding(self.d, flag)
 
+    def set_chunk_size(self, chunk_size: int):
+        _wenet.wenet_set_chunk_size(self.d, chunk_size)
+
     def decode(self,
                audio: Union[str, bytes, np.ndarray],
                last: bool = True) -> str:

runtime/binding/python/wenetruntime/main.py

@@ -15,6 +15,7 @@
 import argparse
 
 from wenetruntime.decoder import Decoder
+from _wenet import wenet_set_log_level as set_log_level  # noqa
 
 
 def get_args():
@@ -23,14 +24,26 @@ def get_args():
                         default='chs',
                         choices=['chs', 'en'],
                         help='select language')
+    parser.add_argument('-c',
+                        '--chunk_size',
+                        default=-1,
+                        type=int,
+                        help='set decoding chunk size')
+    parser.add_argument('-v',
+                        '--verbose',
+                        default=0,
+                        type=int,
+                        help='set log(glog backend) level')
     parser.add_argument('audio', help='input audio file')
     args = parser.parse_args()
     return args
 
 
 def main():
     args = get_args()
+    set_log_level(args.verbose)
     decoder = Decoder(lang=args.language)
+    decoder.set_chunk_size(args.chunk_size)
     result = decoder.decode(args.audio)
     print(result)
 

runtime/core/api/wenet_api.cc

@@ -109,6 +109,8 @@ class Recognizer {
     }
     resource_->post_processor =
         std::make_shared<wenet::PostProcessor>(*post_process_opts_);
+    // Init decode options
+    decode_options_->chunk_size = chunk_size_;
     // Init decoder
     decoder_ = std::make_shared<wenet::AsrDecoder>(feature_pipeline_, resource_,
                                                    *decode_options_);
@@ -180,6 +182,7 @@ class Recognizer {
   void set_context_score(float score) { context_score_ = score; }
   void set_language(const char* lang) { language_ = lang; }
   void set_continuous_decoding(bool flag) { continuous_decoding_ = flag; }
+  void set_chunk_size(int chunk_size) { chunk_size_ = chunk_size; }
 
  private:
   // NOTE(Binbin Zhang): All use shared_ptr for clone in the future
@@ -197,6 +200,7 @@ class Recognizer {
   float context_score_;
   std::string language_ = "chs";
   bool continuous_decoding_ = false;
+  int chunk_size_ = 16;
 };
 
 void* wenet_init(const char* model_dir) {
@@ -255,3 +259,8 @@ void wenet_set_continuous_decoding(void* decoder, int flag) {
   Recognizer* recognizer = reinterpret_cast<Recognizer*>(decoder);
   recognizer->set_continuous_decoding(flag > 0);
 }
+
+void wenet_set_chunk_size(void* decoder, int chunk_size) {
+  Recognizer* recognizer = reinterpret_cast<Recognizer*>(decoder);
+  recognizer->set_chunk_size(chunk_size);
+}

runtime/core/api/wenet_api.h

@@ -103,6 +103,10 @@ void wenet_set_log_level(int level);
  */
 void wenet_set_continuous_decoding(void* decoder, int flag);
 
+/** Set chunk size for decoding, -1 for non-streaming decoding
+ */
+void wenet_set_chunk_size(void* decoder, int chunk_size);
+
 #ifdef __cplusplus
 }
 #endif