fix the asr online client bug, return None, test=doc

paddlespeech/s2t/modules/encoder.py

@@ -317,8 +317,6 @@ class BaseEncoder(nn.Layer):
         outputs = []
         offset = 0
         # Feed forward overlap input step by step
-        print(f"context: {context}")
-        print(f"stride: {stride}")
         for cur in range(0, num_frames - context + 1, stride):
             end = min(cur + decoding_window, num_frames)
             chunk_xs = xs[:, cur:end, :]

paddlespeech/server/README.md

@@ -35,3 +35,16 @@
  ```bash
  paddlespeech_client cls --server_ip 127.0.0.1 --port 8090 --input input.wav
  ```
+
+ ## Online ASR Server
+
+### Lanuch online asr server
+```
+paddlespeech_server start --config_file conf/ws_conformer_application.yaml
+```
+
+### Access online asr server
+
+```
+paddlespeech_client asr_online  --server_ip 127.0.0.1 --port 8090 --input input_16k.wav
+```
\ No newline at end of file

paddlespeech/server/README_cn.md

@@ -35,3 +35,17 @@
  ```bash
  paddlespeech_client cls --server_ip 127.0.0.1 --port 8090 --input input.wav
  ```
+
+## 流式ASR
+
+### 启动流式语音识别服务
+
+```
+paddlespeech_server start --config_file conf/ws_conformer_application.yaml
+```
+
+### 访问流式语音识别服务
+
+```
+paddlespeech_client asr_online  --server_ip 127.0.0.1 --port 8090 --input zh.wav
+```
\ No newline at end of file

paddlespeech/server/bin/paddlespeech_client.py

@@ -277,11 +277,12 @@ class ASRClientExecutor(BaseExecutor):
                 lang=lang,
                 audio_format=audio_format)
             time_end = time.time()
-            logger.info(res.json())
+            logger.info(res)
             logger.info("Response time %f s." % (time_end - time_start))
             return True
         except Exception as e:
             logger.error("Failed to speech recognition.")
+            logger.error(e)
             return False
 
     @stats_wrapper
@@ -299,9 +300,10 @@ class ASRClientExecutor(BaseExecutor):
         logging.info("asr websocket client start")
         handler = ASRAudioHandler(server_ip, port)
         loop = asyncio.get_event_loop()
-        loop.run_until_complete(handler.run(input))
+        res = loop.run_until_complete(handler.run(input))
         logging.info("asr websocket client finished")
 
+        return res['asr_results']
 
 @cli_client_register(
     name='paddlespeech_client.cls', description='visit cls service')

paddlespeech/server/engine/asr/online/asr_engine.py

@@ -473,7 +473,7 @@ class PaddleASRConnectionHanddler:
         ctc_probs = self.model.ctc.log_softmax(ys)  # (1, maxlen, vocab_size)
         ctc_probs = ctc_probs.squeeze(0)
 
-        self.searcher.search(None, ctc_probs, self.cached_feat.place)
+        self.searcher.search(ctc_probs, self.cached_feat.place)
 
         self.hyps = self.searcher.get_one_best_hyps()
         assert self.cached_feat.shape[0] == 1
@@ -823,7 +823,7 @@ class ASRServerExecutor(ASRExecutor):
         ctc_probs = self.model.ctc.log_softmax(
             encoder_out)  # (1, maxlen, vocab_size)
         ctc_probs = ctc_probs.squeeze(0)
-        self.searcher.search(xs, ctc_probs, xs.place)
+        self.searcher.search(ctc_probs, xs.place)
         # update the one best result
         self.hyps = self.searcher.get_one_best_hyps()
 

paddlespeech/server/engine/asr/online/ctc_search.py

@@ -24,19 +24,18 @@ class CTCPrefixBeamSearch:
         """Implement the ctc prefix beam search
 
         Args:
-            config (_type_): _description_
+            config (yacs.config.CfgNode): _description_
         """
         self.config = config
         self.reset()
 
-    def search(self, xs, ctc_probs, device, blank_id=0):
+    def search(self, ctc_probs, device, blank_id=0):
         """ctc prefix beam search method decode a chunk feature
 
         Args:
             xs (paddle.Tensor): feature data
             ctc_probs (paddle.Tensor): the ctc probability of all the tokens
-            encoder_out (paddle.Tensor): _description_
-            encoder_mask (_type_): _description_
+            device (paddle.fluid.core_avx.Place): the feature host device, such as CUDAPlace(0).
             blank_id (int, optional): the blank id in the vocab. Defaults to 0.
 
         Returns:
@@ -45,7 +44,6 @@ class CTCPrefixBeamSearch:
         # decode 
         logger.info("start to ctc prefix search")
 
-        # device = xs.place
         batch_size = 1
         beam_size = self.config.beam_size
         maxlen = ctc_probs.shape[0]

paddlespeech/server/tests/asr/online/websocket_client.py

@@ -34,10 +34,9 @@ class ASRAudioHandler:
     def read_wave(self, wavfile_path: str):
         samples, sample_rate = soundfile.read(wavfile_path, dtype='int16')
         x_len = len(samples)
-        # chunk_stride = 40 * 16  #40ms, sample_rate = 16kHz
-        chunk_size = 80 * 16  #80ms, sample_rate = 16kHz
 
-        if x_len % chunk_size != 0:
+        chunk_size = 85 * 16  #80ms, sample_rate = 16kHz
+        if x_len % chunk_size!= 0:
             padding_len_x = chunk_size - x_len % chunk_size
         else:
             padding_len_x = 0
@@ -48,7 +47,6 @@ class ASRAudioHandler:
         assert (x_len + padding_len_x) % chunk_size == 0
         num_chunk = (x_len + padding_len_x) / chunk_size
         num_chunk = int(num_chunk)
-
         for i in range(0, num_chunk):
             start = i * chunk_size
             end = start + chunk_size
@@ -82,7 +80,6 @@ class ASRAudioHandler:
                 msg = json.loads(msg)
                 logging.info("receive msg={}".format(msg))
 
-            result = msg
             # finished 
             audio_info = json.dumps(
                 {
@@ -98,8 +95,8 @@ class ASRAudioHandler:
             
             # decode the bytes to str
             msg = json.loads(msg)
-            logging.info("receive msg={}".format(msg))
-
+            logging.info("final receive msg={}".format(msg))
+            result = msg
             return result