From fa707ac30c7a66e4888012415590917f6e5cfa34 Mon Sep 17 00:00:00 2001
From: Roman Bazalevsky <rvb@rvb.name>
Date: Sat, 28 Nov 2020 10:50:27 +0300
Subject: [PATCH] =?utf8?q?=D0=93=D0=BE=D0=BB=D0=BE=D1=81=D0=BE=D0=B2=D0=BE?=
 =?utf8?q?=D0=B5=20=D1=83=D0=BF=D1=80=D0=B0=D0=B2=D0=BB=D0=B5=D0=BD=D0=B8?=
 =?utf8?q?=D0=B5:?=
MIME-Version: 1.0
Content-Type: text/plain; charset=utf8
Content-Transfer-Encoding: 8bit

- на сервере устанавливается vosk-api, голосовая модель и запускается vosk-server
- на клиенте устанавливается WebRTC VAD и запускается vosk-client

Из-за некоторых особенностей OpenHAB использует такой протокол:

- при обнаружении ключевой фразы клиент принимает команду и через REST API передает ее исполнителю
- исполнитель обрабатывает команду и по завершении отвечает клиенту на запрос по другому URL
- клиент передает пустую команду, после этого ответ тоже начинает отдаваться пустым

Каждый клиентский хост использует свой URL, соответствующий Item'у в OpenHAB. Сервер, распознающий речь -
общий на локальную сеть.
---
 voicecontrol        | 394 ++++++++++++++++++++++++++++++++++++++++++++
 voicecontrol.ini    |  21 +++
 vosk-client.service |  13 ++
 vosk-server         |  83 ++++++++++
 vosk-server.service |  12 ++
 5 files changed, 523 insertions(+)
 create mode 100755 voicecontrol
 create mode 100644 voicecontrol.ini
 create mode 100644 vosk-client.service
 create mode 100755 vosk-server
 create mode 100644 vosk-server.service

diff --git a/voicecontrol b/voicecontrol
new file mode 100755
index 0000000..fafda2c
--- /dev/null
+++ b/voicecontrol
@@ -0,0 +1,394 @@
+#!/usr/bin/env python3.8
+
+import websockets,asyncio
+import sys
+from pyaudio import PyAudio, Stream, paInt16
+from contextlib import asynccontextmanager, contextmanager, AsyncExitStack
+from typing import AsyncGenerator, Generator
+
+from urllib.parse import urlencode, quote
+import urllib3, base64, json
+
+import configparser 
+from os.path import expanduser
+from streamp3 import MP3Decoder
+
+from time import time, sleep
+
+import webrtcvad
+
+@contextmanager
+def _pyaudio() -> Generator[PyAudio, None, None]:
+    p = PyAudio()
+    try:
+        yield p
+    finally:
+        print('Terminating PyAudio object')
+        p.terminate()
+
+@contextmanager
+def _pyaudio_open_stream(p: PyAudio, *args, **kwargs) -> Generator[Stream, None, None]:
+    s = p.open(*args, **kwargs)
+    try:
+        yield s
+    finally:
+        print('Closing PyAudio Stream')
+        s.close()
+
+@asynccontextmanager
+async def _polite_websocket(ws: websockets.WebSocketClientProtocol) -> AsyncGenerator[websockets.WebSocketClientProtocol, None]:
+    try:
+        yield ws
+    finally:
+        print('Terminating connection')
+        await ws.send('{"eof" : 1}')
+        print(await ws.recv())
+
+def SkipSource(source,seconds):
+  global config
+  try:
+    if config["debug"]:
+     print("Skipping: ", seconds)
+    bufs = int((seconds)*source._rate/source._frames_per_buffer)
+    for i in range(bufs):
+      buffer = source.read(source._frames_per_buffer)
+  except KeyboardInterrupt:
+    raise
+  except:
+    pass
+
+def PlayBack(pyaud, text, mic = None):
+  global config, last_time
+  
+  http = urllib3.PoolManager()
+
+  playback_url = config["tts_url"]
+  playback_param = config["tts_param"]
+
+  if playback_url and text:
+
+    try:
+
+      if playback_param:
+        url = playback_url.format(urlencode({playback_param:text}))
+      else:
+        url = playback_url+quote(text)  
+
+      req = http.request('GET', url, preload_content=False)
+      decoder = MP3Decoder(req)
+
+      speaker = pyaud.open(output=True, format=paInt16, channels=decoder.num_channels, rate=decoder.sample_rate)
+
+      for chunk in decoder:
+        speaker.write(chunk)
+
+      sleep(0.1)
+      speaker.stop_stream()
+      speaker.close()
+
+      elapsed = time() - last_time
+      last_time = time()
+
+      if mic:
+        SkipSource(mic, elapsed + 0.5)
+
+      return elapsed
+
+    except KeyboardInterrupt:
+      raise
+
+    except:
+      pass
+
+  else:
+    return 0
+
+def RunCommand(command, pyaud, mic = None):
+
+  global config
+  
+  http = urllib3.PoolManager()
+
+  command_url = config["command_url"]
+  reply_url = config["reply_url"]
+  command_user = config["api_user"]
+  command_pwd = config["api_pwd"]
+  api_attempts = config["api_attempts"]
+
+  if command_url:
+    try:
+      if command_user:
+        my_headers = urllib3.util.make_headers(basic_auth=command_user+':'+command_pwd)
+      else:
+        my_headers = urllib3.util.make_headers()  
+      my_headers['Content-Type']='text/plain'
+      my_headers['Accept']='apllication/json'
+      http.request('POST',command_url,headers=my_headers,body=command.encode('UTF-8'))
+      if reply_url:
+        sleep(0.5)
+        res="NULL"
+        for i in range(api_attempts):
+          try:
+            if command_user:
+              my_headers = urllib3.util.make_headers(basic_auth=command_user+':'+command_pwd)
+            else:
+              my_headers = urllib3.util.make_headers()  
+            req=http.request('GET',reply_url,headers=my_headers).data
+            res = json.loads(req)['state'].strip()
+            if config["debug"]:
+              print(res)
+            if not(res == 'NULL'):
+              break
+            sleep(1)  
+          except KeyboardInterrupt:
+            raise
+          except:  
+            sleep(1)
+        if res and not(res=="NULL"):
+          PlayBack(pyaud, res, mic=mic)
+        elif res=="NULL":
+          PlayBack(pyaud, "Сервер не ответил", mic=mic)  
+      http.request('POST',command_url, headers=my_headers, body="")
+    except KeyboardInterrupt:
+      raise
+    except:
+      try:
+        http.request('POST',command_url, headers=my_headers, body="")
+      except:  
+        pass
+
+async def ListenPhrase(mic, server):
+  global config,last_time, vad
+
+  frame = 30/1000 # 30 ms
+  pause = 2
+  sz = int(mic._rate*frame)
+  sp = int(pause/frame)
+
+  try:
+
+    phrase = ""
+    voice = False
+
+    while not phrase:
+      data = mic.read(sz)
+      if len(data) == 0:
+        break
+      vd = vad.is_speech(data, mic._rate)
+      if vd and not voice:
+        voice = True
+        if config["debug"]:
+          print("+", end="")
+        cnt = 0
+      if voice and not vd:
+        cnt = cnt + 1
+        if cnt > sp:
+          cnt = 0
+          voice = False
+          if config["debug"]:
+            print("-")
+      if voice:
+        print("*",end="")
+        await server.send(data)
+        datatxt = await server.recv()
+        data = json.loads(datatxt)
+        try:
+          phrase = data["text"]
+          confidence = min(map(lambda x: x["conf"], data["result"]))
+        except:
+          pass  
+  
+    last_time = time()
+
+    return phrase, confidence
+
+  except KeyboardInterrupt:
+    raise
+  except websockets.exceptions.ConnectionClosedError:
+    raise  
+  except:
+    raise
+    return '',0
+
+
+async def hello(uri):
+
+  global config, last_time
+
+  keyphrase = config["keyphrase"]
+  confidence_treshold = config["confidence_treshold"]
+  rec_attempts = config["rec_attempts"]
+  commands = config["commands"]
+
+  async with AsyncExitStack() as stack:
+    ws = await stack.enter_async_context(websockets.connect(uri))
+    print('Type Ctrl-C to exit')
+    phrases = config["commands"]
+    phrases.append(config["keyphrase"])
+    phrases = json.dumps(phrases, ensure_ascii=False)
+    await ws.send('{"config" : { "phrase_list" : '+phrases+', "sample_rate" : 16000.0}}')
+
+    ws = await stack.enter_async_context(_polite_websocket(ws))
+    p = stack.enter_context(_pyaudio())
+    s = stack.enter_context(_pyaudio_open_stream(p,
+            format = paInt16, 
+            channels = 1,
+            rate = 16000,
+            input = True, 
+            frames_per_buffer = 2000))
+    while True:
+      phrase, confidence = await ListenPhrase(s, ws)
+      if config["debug"]:
+        print(phrase,confidence)
+      if phrase == keyphrase and confidence>=confidence_treshold :
+        PlayBack(p, "Я жду команду", mic=s)
+        command = ""
+
+        for i in range(rec_attempts):
+          phrase, confidence = await ListenPhrase(s, ws)
+          if config["debug"]:
+            print(phrase,confidence)
+          if confidence > confidence_treshold:
+            if (not commands) or (phrase in commands):
+              if config["debug"]:
+                print("Command: ", phrase)
+              command = phrase
+              RunCommand(command, p, s)
+              break
+            else:
+              PlayBack(p, "Не знаю такой команды: "+phrase, mic=s)
+          else:
+            PlayBack(p, "Не поняла, слишком неразборчиво", mic=s)
+
+        if not command:
+          PlayBack(p, "Так команду и не поняла...", mic=s)
+
+             
+def get_config(path):
+  
+  config = configparser.ConfigParser()
+  config.read(path)
+  
+  try:  
+    keyphrase = config['vosk']['keyphrase']
+  except:
+    print ("Обязательный параметр - ключевое слово - не задан!")
+    raise
+
+  try:  
+    rec_attempts = int(config['vosk']['attempts'])
+  except:
+    rec_attempts = 4
+
+  try:  
+    confidence_treshold = float(config['vosk']['confidence_treshold'])
+  except:
+    confidence_treshold = 0.4
+
+  try:
+    vosk_server = config['vosk']['server']
+  except:
+    print ("Обязательный параметр - сервер распознавания - не задан!")
+    raise
+
+  try:
+    command_file=config['commands']['command_file']
+    with open(command_file) as file:
+      commands = file.read().splitlines()
+  except:
+    commands = None
+
+  try:
+    tts_url=config['rest']['tts_url']
+  except:
+    tts_url = None
+
+  try:
+    tts_param=config['rest']['tts_param']
+  except:
+    tts_param = None
+
+  try:
+    api_attempts=int(config['rest']['attempts'])
+  except:
+    api_attempts = 2
+
+  try:  
+    api_user=config['rest']['api_user']
+    api_pwd=config['rest']['api_pwd']
+  except:
+    api_user = None
+    api_pwd = None
+
+  try:
+    command_url=config['rest']['command_url']
+  except:
+    command_url = None
+
+  try:  
+    reply_url=config['rest']['reply_url']
+  except:
+    reply_url = None  
+
+  try:  
+    vad_mode=config['vad']['agressive']
+  except:
+    vad_mode = 3
+
+  try:
+    debug = (config['system']['debug'].lower() == "true")
+  except:
+    debug = False  
+
+  if command_file:
+    with open(command_file) as file:
+      commands = file.read().splitlines()
+
+  return {
+      "asr_server": vosk_server,
+      "keyphrase": keyphrase,
+      "rec_attempts": rec_attempts,
+      "confidence_treshold": confidence_treshold,
+      "tts_url": tts_url,
+      "tts_param": tts_param,
+      "api_attempts": api_attempts,
+      "api_user": api_user,
+      "api_pwd": api_pwd,
+      "command_url": command_url,
+      "reply_url": reply_url,
+      "debug": debug,
+      "commands": commands,
+      "vad_mode": vad_mode
+    }
+
+
+if len(sys.argv) == 2:
+    conf_file = sys.argv[1]
+else:
+    conf_file = expanduser("~")+"/.config/voicecontrol.ini"
+
+config = get_config(conf_file)
+
+server = config['asr_server']
+
+vad = webrtcvad.Vad(config['vad_mode'])
+
+while True:
+
+  try:
+
+    loop = asyncio.get_event_loop()
+    loop.run_until_complete(
+        hello(f'ws://' + server))
+
+  except (Exception, KeyboardInterrupt) as e:
+    loop.run_until_complete(
+      loop.shutdown_asyncgens())
+    if isinstance(e, KeyboardInterrupt):
+      loop.stop()
+      print('Bye')
+      exit(0)
+    else:
+      print(f'Oops! {e}')
+      print('Restarting process...')
+      sleep(10)
diff --git a/voicecontrol.ini b/voicecontrol.ini
new file mode 100644
index 0000000..bceffb3
--- /dev/null
+++ b/voicecontrol.ini
@@ -0,0 +1,21 @@
+[vosk]
+keyphrase = окей гестия
+attempts = 3
+confidence_threshold = 0.6
+server = 192.168.1.100:2700
+
+[commands]
+command_file = /etc/vosk/commands.txt
+
+[rest]
+attempts = 5
+tts_url = http://estia.rvb-home.lan/festival?{0}
+tts_param = q
+
+api_user = openhabrest
+api_pwd = somestrictpwd
+command_url = https://openhab.rvb.name/rest/items/Command
+reply_url = https://openhab.rvb.name/rest/items/Reply
+
+[system]
+debug = True
diff --git a/vosk-client.service b/vosk-client.service
new file mode 100644
index 0000000..5bfb7b2
--- /dev/null
+++ b/vosk-client.service
@@ -0,0 +1,13 @@
+[Unit]
+Description=Vosk ASR service client
+
+[Service]
+Environment=PULSE_SERVER=localhost
+Type=simple
+ExecStartPre=/usr/bin/pactl set-source-volume alsa_input.usb-TP6920_TP6920_0621-00.analog-mono 135%
+ExecStart=/usr/local/bin/voicecontrol /etc/vosk/voicecontrol.ini
+Restart=always
+RestartSec=30s
+
+[Install]
+WantedBy=multi-user.target
diff --git a/vosk-server b/vosk-server
new file mode 100755
index 0000000..78f1210
--- /dev/null
+++ b/vosk-server
@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+
+import json
+import os
+import sys
+import asyncio
+import pathlib
+import websockets
+import concurrent.futures
+import logging
+from vosk import Model, KaldiRecognizer
+
+from pprint import pprint
+
+# Enable loging if needed
+#
+# logger = logging.getLogger('websockets')
+# logger.setLevel(logging.INFO)
+# logger.addHandler(logging.StreamHandler())
+
+vosk_interface = os.environ.get('VOSK_SERVER_INTERFACE', '0.0.0.0')
+vosk_port = int(os.environ.get('VOSK_SERVER_PORT', 2700))
+vosk_model_path = os.environ.get('VOSK_MODEL_PATH', 'model')
+vosk_sample_rate = float(os.environ.get('VOSK_SAMPLE_RATE', 8000))
+
+if len(sys.argv) > 1:
+   vosk_model_path = sys.argv[1]
+
+# Gpu part, uncomment if vosk-api has gpu support
+#
+# from vosk import GpuInit, GpuInstantiate
+# GpuInit()
+# def thread_init():
+#     GpuInstantiate()
+# pool = concurrent.futures.ThreadPoolExecutor(initializer=thread_init)
+
+model = Model(vosk_model_path)
+pool = concurrent.futures.ThreadPoolExecutor((os.cpu_count() or 1))
+loop = asyncio.get_event_loop()
+
+def process_chunk(rec, message):
+    if message == '{"eof" : 1}':
+        return rec.FinalResult(), True
+    elif rec.AcceptWaveform(message):
+        return rec.Result(), False
+    else:
+        return rec.PartialResult(), False
+
+async def recognize(websocket, path):
+
+    rec = None
+    phrase_list = None
+    sample_rate = vosk_sample_rate
+
+    while True:
+
+        message = await websocket.recv()
+
+        # Load configuration if provided
+        if isinstance(message, str) and 'config' in message:
+            jobj = json.loads(message)['config']
+            if 'phrase_list' in jobj:
+                phrase_list = jobj['phrase_list']
+            if 'sample_rate' in jobj:
+                sample_rate = float(jobj['sample_rate'])
+            continue
+
+        # Create the recognizer, word list is temporary disabled since not every model supports it
+        if not rec:
+            if phrase_list:
+                 rec = KaldiRecognizer(model, sample_rate, json.dumps(phrase_list, ensure_ascii=False))
+            else:
+                 rec = KaldiRecognizer(model, sample_rate)
+
+        response, stop = await loop.run_in_executor(pool, process_chunk, rec, message)
+        await websocket.send(response)
+        if stop: break
+
+start_server = websockets.serve(
+    recognize, vosk_interface, vosk_port)
+
+loop.run_until_complete(start_server)
+loop.run_forever()
diff --git a/vosk-server.service b/vosk-server.service
new file mode 100644
index 0000000..a8957d2
--- /dev/null
+++ b/vosk-server.service
@@ -0,0 +1,12 @@
+[Unit]
+Description=Vosk ASR service
+
+[Service]
+Environment=VOSK_MODEL_PATH=/opt/vosk/vosk-model-ru-0.10-lgraph
+Type=simple
+ExecStart=/usr/local/bin/vosk-server
+Restart=always
+RestartSec=30s
+
+[Install]
+WantedBy=multi-user.target
-- 
2.34.1