[ros_speech_recognition] Add vosk engine

ros_speech_recognition/CMakeLists.txt

@@ -12,17 +12,14 @@ generate_dynamic_reconfigure_options(
   cfg/SpeechRecognition.cfg
 )
 
+add_custom_target(${PROJECT_NAME}_install_trained_data ALL COMMAND python$ENV{ROS_PYTHON_VERSION} ${PROJECT_SOURCE_DIR}/scripts/install_trained_data.py)
+
 catkin_package()
 
-if($ENV{ROS_DISTRO} STRGREATER "melodic")
-  catkin_generate_virtualenv(
-    PYTHON_INTERPRETER python3
-    )
-else()
-  catkin_generate_virtualenv(
-    PYTHON_INTERPRETER python2
+catkin_generate_virtualenv(
+  PYTHON_INTERPRETER python3
+  CHECK_VENV FALSE
     )
-endif()
 
 file(GLOB PYTHON_SCRIPT_FILES scripts/*.py test/*.py)
 catkin_install_python(

ros_speech_recognition/README.md

@@ -220,6 +220,14 @@ roslaunch ros_speech_recognition parrotry.launch language:=ja-JP
 
   Auth key for Bing API.  
   This is valid only if `~engine` is `bing`.
+
+* `~vosk_model_path` (`String`, default: `None`)
+
+  Path to trainded model for Vosk API.
+  This is valid only if `~engine` is `Vosk`.
+
+  If `en-US` or `ja` is selected as `~language`, you do not need to specify the path.
+  To load other models, please download them from [Model list](https://alphacephei.com/vosk/models).
 
 ## Author
 

ros_speech_recognition/cfg/SpeechRecognition.cfg

@@ -10,7 +10,8 @@ engine_enum = gen.enum([gen.const("Google", str_t, "Google", "Google Speech Reco
                         gen.const("Wit", str_t, "Wit", "Wit.ai API"),
                         gen.const("Bing", str_t, "Bing", "Microsoft Bing Speech API"),
                         gen.const("Houndify", str_t, "Houndify", "Houndify API"),
-                        gen.const("IBM", str_t, "IBM", "IBM Speech to Text API")],
+                        gen.const("IBM", str_t, "IBM", "IBM Speech to Text API"),
+                        gen.const("Vosk", str_t, "Vosk", "Vosk API")],
                        "engine")
 
 #       name    type     level     description     default      min      max

ros_speech_recognition/package.xml

@@ -13,13 +13,15 @@
 
   <build_depend>catkin_virtualenv</build_depend>
   <build_depend>dynamic_reconfigure</build_depend>
+  <build_depend>jsk_data</build_depend>
   <build_depend>speech_recognition_msgs</build_depend>
   <build_depend>g++-static</build_depend>
 
   <run_depend>audio_capture</run_depend>
   <run_depend>audio_common_msgs</run_depend>
   <run_depend>dynamic_reconfigure</run_depend>
   <run_depend>flac</run_depend>
+  <run_depend>jsk_data</run_depend>
   <run_depend>sound_play</run_depend>
   <run_depend>speech_recognition_msgs</run_depend>
   <run_depend>ubuntu-sounds</run_depend>

ros_speech_recognition/requirements.txt

@@ -1 +1,2 @@
-SpeechRecognition==3.8.1
+SpeechRecognition==3.9.0
+vosk==0.3.45

ros_speech_recognition/scripts/install_trained_data.py

@@ -0,0 +1,44 @@
+#!/usr/bin/env python
+
+import argparse
+import multiprocessing
+
+import jsk_data
+
+
+def download_data(*args, **kwargs):
+    p = multiprocessing.Process(
+            target=jsk_data.download_data,
+            args=args,
+            kwargs=kwargs)
+    p.start()
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-v', '--verbose', dest='quiet', action='store_false')
+    args = parser.parse_args()
+    quiet = args.quiet
+
+    PKG = 'ros_speech_recognition'
+
+    download_data(
+        pkg_name=PKG,
+        path='trained_data/vosk-model-small-ja-0.22.zip',
+        url='https://alphacephei.com/vosk/models/vosk-model-small-ja-0.22.zip',  # NOQA
+        md5='0e3163dd62dfb0d823353718ac3cbf79',
+        extract=True,
+        quiet=quiet,
+    )
+
+    download_data(
+        pkg_name=PKG,
+        path='trained_data/vosk-model-small-en-us-0.15.zip',
+        url='https://alphacephei.com/vosk/models/vosk-model-small-en-us-0.15.zip',  # NOQA
+        md5='09ab50ccd62b674cbaa231b825f9c1cb',
+        extract=True,
+        quiet=quiet,
+    )
+
+if __name__ == '__main__':
+    main()

ros_speech_recognition/scripts/speech_recognition_node.py

@@ -6,6 +6,7 @@
 import rospy
 import speech_recognition as SR
 from ros_speech_recognition.recognize_google_cloud import RecognizerEx
+import ros_speech_recognition.recognize_vosk
 import json
 import array
 import sys
@@ -275,6 +276,10 @@ def recognize(self, audio):
             recog_func = self.recognizer.recognize_houndify
         elif self.engine == Config.SpeechRecognition_IBM:
             recog_func = self.recognizer.recognize_ibm
+        elif self.engine == Config.SpeechRecognition_Vosk:
+            if not self.args:
+                self.args = {'model_path': rospy.get_param('~vosk_model_path', None)}
+            recog_func = self.recognizer.recognize_vosk
 
         return recog_func(audio_data=audio, language=self.language, **self.args)
 

ros_speech_recognition/src/ros_speech_recognition/recognize_vosk.py

@@ -0,0 +1,37 @@
+# file to override recognize_vosk
+# we need this to use vosk model anywhere
+
+from speech_recognition import AudioData
+from ros_speech_recognition.recognize_google_cloud import RecognizerEx
+from vosk import Model, KaldiRecognizer
+import json
+import os.path as osp
+import rospkg
+import rospy
+
+def recognize_vosk(self, audio_data, model_path=None, language='en-US'):
+
+    assert isinstance(audio_data, AudioData), "Data must be audio data"
+
+    if not hasattr(self, 'vosk_model'):
+        if model_path is None:
+            PKG = 'ros_speech_recognition'
+            rp = rospkg.RosPack()
+            data_path = osp.join(rp.get_path(PKG), 'trained_data')
+            if language == 'en-US':
+                model_path = osp.join(data_path, 'vosk-model-small-en-us-0.15')
+            elif language == 'ja':
+                model_path = osp.join(data_path, 'vosk-model-small-ja-0.22')
+            else:
+                rospy.logerr("Unsupported language: {0}.\n Please download the model from https://alphacephei.com/vosk/models and specify its path as 'vosk_model_path'.".format(language))
+                exit (1)
+        rospy.loginfo("Loading model from {}".format(model_path))
+        self.vosk_model = Model(model_path)
+    rec = KaldiRecognizer(self.vosk_model, 16000);
+
+    rec.AcceptWaveform(audio_data.get_raw_data(convert_rate=16000, convert_width=2));
+    finalRecognition = rec.FinalResult()
+    text = json.loads(finalRecognition)['text']
+    return text
+
+RecognizerEx.recognize_vosk = recognize_vosk

ros_speech_recognition/trained_data/.gitignore

@@ -0,0 +1,2 @@
+*
+!.gitignore