Merge pull request #467 from alibaba-damo-academy/dev_websocket_offline

websocket offline

funasr/runtime/python/websocket/README.md

@@ -22,15 +22,13 @@ pip install -r requirements_server.txt
 
 ### Start server
 #### ASR offline server
+```shell
+python ws_server_offline.py --port 10095 --asr_model "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
+```
 
-[//]: # (```shell)
-
-[//]: # (python ws_server_online.py --host "0.0.0.0" --port 10095 --asr_model "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch")
-
-[//]: # (```)
 #### ASR streaming server
 ```shell
-python ws_server_online.py --host "0.0.0.0" --port 10095 --asr_model_online "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online"
+python ws_server_online.py --port 10095 --asr_model_online "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online"
 ```
 
 #### ASR offline/online 2pass server
@@ -51,17 +49,31 @@ pip install -r requirements_client.txt
 ```
 
 ### Start client
-#### Recording from mircrophone
+#### ASR offline client
+##### Recording from mircrophone
+```shell
+# --chunk_interval, "10": 600/10=60ms, "5"=600/5=120ms, "20": 600/12=30ms
+python ws_client.py --host "0.0.0.0" --port 10095 --chunk_interval 10 --words_max_print 100
+```
+##### Loadding from wav.scp(kaldi style)
+```shell
+# --chunk_interval, "10": 600/10=60ms, "5"=600/5=120ms, "20": 600/12=30ms
+python ws_client.py --host "0.0.0.0" --port 10095 --chunk_interval 10 --words_max_print 100 --audio_in "./data/wav.scp" --send_without_sleep
+```
+#### ASR streaming client
+##### Recording from mircrophone
 ```shell
 # --chunk_size, "5,10,5"=600ms, "8,8,4"=480ms
-python ws_client.py --host "127.0.0.1" --port 10095 --chunk_size "5,10,5" --words_max_print 100
+python ws_client.py --host "0.0.0.0" --port 10095 --chunk_size "5,10,5" --words_max_print 100
 ```
-#### Loadding from wav.scp(kaldi style)
+##### Loadding from wav.scp(kaldi style)
 ```shell
 # --chunk_size, "5,10,5"=600ms, "8,8,4"=480ms
-python ws_client.py --host "127.0.0.1" --port 10095 --chunk_size "5,10,5" --audio_in "./data/wav.scp" --words_max_print 100
+python ws_client.py --host "0.0.0.0" --port 10095 --chunk_size "5,10,5" --audio_in "./data/wav.scp" --words_max_print 100
 ```
 
+#### ASR offline/online 2pass client
+
 ## Acknowledge
 1. This project is maintained by [FunASR community](https://github.com/alibaba-damo-academy/FunASR).
 2. We acknowledge [zhaoming](https://github.com/zhaomingwork/FunASR/tree/fix_bug_for_python_websocket) for contributing the websocket service.

funasr/runtime/python/websocket/parse_args.py

@@ -31,5 +31,10 @@ parser.add_argument("--ngpu",
                     type=int,
                     default=1,
                     help="0 for cpu, 1 for gpu")
+parser.add_argument("--ncpu",
+                    type=int,
+                    default=1,
+                    help="cpu cores")
 
-args = parser.parse_args()
+args = parser.parse_args()
+print(args)

funasr/runtime/python/websocket/ws_client.py

@@ -31,7 +31,10 @@ parser.add_argument("--audio_in",
                     type=str,
                     default=None,
                     help="audio_in")
-
+parser.add_argument("--send_without_sleep",
+                    action="store_true",
+                    default=False,
+                    help="if audio_in is set, send_without_sleep")
 parser.add_argument("--test_thread_num",
                     type=int,
                     default=1,
@@ -43,12 +46,11 @@ parser.add_argument("--words_max_print",
 
 args = parser.parse_args()
 args.chunk_size = [int(x) for x in args.chunk_size.split(",")]
-
+print(args)
 # voices = asyncio.Queue()
 from queue import Queue
 voices = Queue()
 
-# 其他函数可以通过调用send(data)来发送数据，例如：
 async def record_microphone():
     is_finished = False
     import pyaudio
@@ -75,11 +77,9 @@ async def record_microphone():
         message = json.dumps({"chunk_size": args.chunk_size, "chunk_interval": args.chunk_interval, "audio": data, "is_speaking": is_speaking, "is_finished": is_finished})
         
         voices.put(message)
-        #print(voices.qsize())
 
         await asyncio.sleep(0.005)
 
-# 其他函数可以通过调用send(data)来发送数据，例如：
 async def record_from_scp():
     import wave
     global voices
@@ -95,15 +95,11 @@ async def record_from_scp():
         # bytes_f = open(wav_path, "rb")
         # bytes_data = bytes_f.read()
         with wave.open(wav_path, "rb") as wav_file:
-            # 获取音频参数
             params = wav_file.getparams()
-            # 获取头信息的长度
             # header_length = wav_file.getheaders()[0][1]
-            # 读取音频帧数据，跳过头信息
             # wav_file.setpos(header_length)
             frames = wav_file.readframes(wav_file.getnframes())
 
-        # 将音频帧数据转换为字节类型的数据
         audio_bytes = bytes(frames)
         # stride = int(args.chunk_size/1000*16000*2)
         stride = int(60*args.chunk_size[1]/args.chunk_interval/1000*16000*2)
@@ -120,8 +116,8 @@ async def record_from_scp():
             voices.put(message)
             # print("data_chunk: ", len(data_chunk))
             # print(voices.qsize())
-        
-            await asyncio.sleep(60*args.chunk_size[1]/args.chunk_interval/1000)
+            sleep_duration = 0.001 if args.send_without_sleep else 60*args.chunk_size[1]/args.chunk_interval/1000
+            await asyncio.sleep(sleep_duration)
 
     is_finished = True
     message = json.dumps({"is_finished": is_finished})
@@ -136,7 +132,7 @@ async def ws_send():
             data = voices.get()
             voices.task_done()
             try:
-                await websocket.send(data) # 通过ws对象发送数据
+                await websocket.send(data)
             except Exception as e:
                 print('Exception occurred:', e)
                 traceback.print_exc()
@@ -155,9 +151,14 @@ async def message(id):
             meg = json.loads(meg)
             # print(meg, end = '')
             # print("\r")
-            text_print += " {}".format(meg["text"][0]) 
+            # print(meg)
+            text = meg["text"][0]
+            if meg["mode"] == "online":
+                text_print += " {}".format(text)
+            else:
+                text_print += "{}".format(text)
             text_print = text_print[-args.words_max_print:]
-            #os.system('clear')
+            os.system('clear')
             print("\r"+str(id)+":"+text_print)
         except Exception as e:
             print("Exception:", e)
@@ -177,17 +178,15 @@ async def print_messge():
             exit(0)
 
 async def ws_client(id):
-    global websocket # 定义一个全局变量ws，用于保存websocket连接对象
-    # uri = "ws://11.167.134.197:8899"
+    global websocket
     uri = "ws://{}:{}".format(args.host, args.port)
-    #ws = await websockets.connect(uri, subprotocols=["binary"]) # 创建一个长连接
     async for websocket in websockets.connect(uri, subprotocols=["binary"], ping_interval=None):
         if args.audio_in is not None:
-            task = asyncio.create_task(record_from_scp()) # 创建一个后台任务录音
+            task = asyncio.create_task(record_from_scp())
         else:
-            task = asyncio.create_task(record_microphone())  # 创建一个后台任务录音
-        task2 = asyncio.create_task(ws_send()) # 创建一个后台任务发送
-        task3 = asyncio.create_task(message(id)) # 创建一个后台接收消息的任务
+            task = asyncio.create_task(record_microphone())
+        task2 = asyncio.create_task(ws_send())
+        task3 = asyncio.create_task(message(id))
         await asyncio.gather(task, task2, task3)
 
 def one_thread(id):
@@ -198,13 +197,13 @@ def one_thread(id):
 if __name__ == '__main__':
     process_list = []
     for i in range(args.test_thread_num):   
-        p = Process(target=one_thread,args=(i,)) #实例化进程对象
+        p = Process(target=one_thread,args=(i,))
         p.start()
         process_list.append(p)
 
     for i in process_list:
         p.join()
 
-    print('结束测试')
+    print('end')
  
 

funasr/runtime/python/websocket/ws_server_offline.py

@@ -0,0 +1,147 @@
+import asyncio
+import json
+import websockets
+import time
+import logging
+import tracemalloc
+import numpy as np
+
+from parse_args import args
+from modelscope.pipelines import pipeline
+from modelscope.utils.constant import Tasks
+from modelscope.utils.logger import get_logger
+from funasr.runtime.python.onnxruntime.funasr_onnx.utils.frontend import load_bytes
+
+tracemalloc.start()
+
+logger = get_logger(log_level=logging.CRITICAL)
+logger.setLevel(logging.CRITICAL)
+
+
+websocket_users = set()
+
+print("model loading")
+# asr
+inference_pipeline_asr = pipeline(
+    task=Tasks.auto_speech_recognition,
+    model=args.asr_model,
+    ngpu=args.ngpu,
+    ncpu=args.ncpu,
+    model_revision=None)
+
+
+# vad
+inference_pipeline_vad = pipeline(
+    task=Tasks.voice_activity_detection,
+    model=args.vad_model,
+    model_revision=None,
+    output_dir=None,
+    batch_size=1,
+    mode='online',
+    ngpu=args.ngpu,
+    ncpu=args.ncpu,
+)
+
+if args.punc_model != "":
+    inference_pipeline_punc = pipeline(
+        task=Tasks.punctuation,
+        model=args.punc_model,
+        model_revision=None,
+        ngpu=args.ngpu,
+        ncpu=args.ncpu,
+    )
+else:
+    inference_pipeline_punc = None
+
+print("model loaded")
+
+async def ws_serve(websocket, path):
+    frames = []
+    frames_asr = []
+    global websocket_users
+    websocket_users.add(websocket)
+    websocket.param_dict_asr = {}
+    websocket.param_dict_vad = {'in_cache': dict(), "is_final": False}
+    websocket.param_dict_punc = {'cache': list()}
+    websocket.vad_pre_idx = 0
+    speech_start = False
+
+    try:
+        async for message in websocket:
+            message = json.loads(message)
+            is_finished = message["is_finished"]
+            if not is_finished:
+                audio = bytes(message['audio'], 'ISO-8859-1')
+                frames.append(audio)
+                duration_ms = len(audio)//32
+                websocket.vad_pre_idx += duration_ms
+
+                is_speaking = message["is_speaking"]
+                websocket.param_dict_vad["is_final"] = not is_speaking
+                if speech_start:
+                    frames_asr.append(audio)
+                speech_start_i, speech_end_i = await async_vad(websocket, audio)
+                if speech_start_i:
+                    speech_start = True
+                    beg_bias = (websocket.vad_pre_idx-speech_start_i)//duration_ms
+                    frames_pre = frames[-beg_bias:]
+                    frames_asr = []
+                    frames_asr.extend(frames_pre)
+                if speech_end_i or not is_speaking:
+                    audio_in = b"".join(frames_asr)
+                    await async_asr(websocket, audio_in)
+                    frames_asr = []
+                    speech_start = False
+                    if not is_speaking:
+                        websocket.vad_pre_idx = 0
+                        frames = []
+                    else:
+                        frames = frames[-10:]
+
+     
+    except websockets.ConnectionClosed:
+        print("ConnectionClosed...", websocket_users)
+        websocket_users.remove(websocket)
+    except websockets.InvalidState:
+        print("InvalidState...")
+    except Exception as e:
+        print("Exception:", e)
+
+
+async def async_vad(websocket, audio_in):
+
+    segments_result = inference_pipeline_vad(audio_in=audio_in, param_dict=websocket.param_dict_vad)
+
+    speech_start = False
+    speech_end = False
+    
+    if len(segments_result) == 0 or len(segments_result["text"]) > 1:
+        return speech_start, speech_end
+    if segments_result["text"][0][0] != -1:
+        speech_start = segments_result["text"][0][0]
+    if segments_result["text"][0][1] != -1:
+        speech_end = True
+    return speech_start, speech_end
+
+
+async def async_asr(websocket, audio_in):
+            if len(audio_in) > 0:
+                # print(len(audio_in))
+                audio_in = load_bytes(audio_in)
+                
+                rec_result = inference_pipeline_asr(audio_in=audio_in,
+                                                    param_dict=websocket.param_dict_asr)
+                # print(rec_result)
+                if inference_pipeline_punc is not None and 'text' in rec_result and len(rec_result["text"])>0:
+                    rec_result = inference_pipeline_punc(text_in=rec_result['text'],
+                                                         param_dict=websocket.param_dict_punc)
+                    # print(rec_result)
+                    message = json.dumps({"mode": "offline", "text": [rec_result["text"]]})
+                    await websocket.send(message)
+                    
+ 
+
+
+start_server = websockets.serve(ws_serve, args.host, args.port, subprotocols=["binary"], ping_interval=None)
+asyncio.get_event_loop().run_until_complete(start_server)
+asyncio.get_event_loop().run_forever()

funasr/runtime/python/websocket/ws_server_online.py

@@ -12,7 +12,7 @@ from parse_args import args
 from modelscope.pipelines import pipeline
 from modelscope.utils.constant import Tasks
 from modelscope.utils.logger import get_logger
-from funasr_onnx.utils.frontend import load_bytes
+from funasr.runtime.python.onnxruntime.funasr_onnx.utils.frontend import load_bytes
 
 tracemalloc.start()
 
@@ -28,6 +28,8 @@ print("model loading")
 inference_pipeline_asr_online = pipeline(
     task=Tasks.auto_speech_recognition,
     model=args.asr_model_online,
+    ngpu=args.ngpu,
+    ncpu=args.ncpu,
     model_revision='v1.0.4')
 
 print("model loaded")
@@ -63,14 +65,14 @@ async def ws_serve(websocket, path):
 
      
     except websockets.ConnectionClosed:
-        print("ConnectionClosed...", websocket_users)    # 链接断开
+        print("ConnectionClosed...", websocket_users)
         websocket_users.remove(websocket)
     except websockets.InvalidState:
-        print("InvalidState...")    # 无效状态
+        print("InvalidState...")
     except Exception as e:
         print("Exception:", e)
  
-async def async_asr_online(websocket,audio_in): # ASR推理
+async def async_asr_online(websocket,audio_in):
             if len(audio_in) > 0:
                 audio_in = load_bytes(audio_in)
                 rec_result = inference_pipeline_asr_online(audio_in=audio_in,
@@ -84,7 +86,6 @@ async def async_asr_online(websocket,audio_in): # ASR推理
                         message = json.dumps({"mode": "online", "text": rec_result["text"]})
                         await websocket.send(message)
 
- 
 
 
 start_server = websockets.serve(ws_serve, args.host, args.port, subprotocols=["binary"], ping_interval=None)