[assistant.picovoice] Conversation flow improvements.

- The `Responding` state should be modelled as an extra event/binary
  flag, not as an assistant state. The assistant may be listening for
  hotwords even while the `tts` plugin is responding, and we don't want
  the two states to interfere with each either - neither to build a more
  complex state machine that also needs to take concurrent states into
  account.

- Stop any responses being rendered upon the `tts` plugin when a new
  hotword audio is detected. If e.g. I say "Ok Google", I should always
  be able to trigger the assistant and stop any concurrent audio
  process.

- `SpeechRecognizedEvent` should be emitted even if `cheetah`'s latest
  audio frame results weren't marked as final, and the speech detection
  window timed out. Cheetah's `is_final` detection seems to be quite
  buggy sometimes, and it may not properly detect the end of utterances,
  especially with non-native accents. The workaround is to flush out
  whatever text is available (if at least some speech was detected) into
  a `SpeechRecognizedEvent` upon timeout.

platypush/__init__.py

@@ -7,10 +7,13 @@ Platypush
 
 from .app import Application
 from .config import Config
-from .context import get_backend, get_bus, get_plugin
+from .context import Variable, get_backend, get_bus, get_plugin
+from .cron import cron
+from .event.hook import hook
 from .message.event import Event
 from .message.request import Request
 from .message.response import Response
+from .procedure import procedure
 from .runner import main
 from .utils import run
 
@@ -19,14 +22,18 @@ __author__ = 'Fabio Manganiello <fabio@manganiello.tech>'
 __version__ = '0.50.3'
 __all__ = [
     'Application',
+    'Variable',
     'Config',
     'Event',
     'Request',
     'Response',
+    'cron',
     'get_backend',
     'get_bus',
     'get_plugin',
+    'hook',
     'main',
+    'procedure',
     'run',
 ]
 

platypush/message/event/__init__.py

@@ -257,26 +257,29 @@ class Event(Message):
 
         return result
 
+    def as_dict(self):
+        """
+        Converts the event into a dictionary
+        """
+        args = copy.deepcopy(self.args)
+        flatten(args)
+        return {
+            'type': 'event',
+            'target': self.target,
+            'origin': self.origin if hasattr(self, 'origin') else None,
+            'id': self.id if hasattr(self, 'id') else None,
+            '_timestamp': self.timestamp,
+            'args': {'type': self.type, **args},
+        }
+
     def __str__(self):
         """
         Overrides the str() operator and converts
         the message into a UTF-8 JSON string
         """
-
         args = copy.deepcopy(self.args)
         flatten(args)
-
-        return json.dumps(
-            {
-                'type': 'event',
-                'target': self.target,
-                'origin': self.origin if hasattr(self, 'origin') else None,
-                'id': self.id if hasattr(self, 'id') else None,
-                '_timestamp': self.timestamp,
-                'args': {'type': self.type, **args},
-            },
-            cls=self.Encoder,
-        )
+        return json.dumps(self.as_dict(), cls=self.Encoder)
 
 
 @dataclass

platypush/message/event/assistant/__init__.py

@@ -1,27 +1,53 @@
 import re
 import sys
-from typing import Optional
+from typing import Optional, Union
 
 from platypush.context import get_plugin
 from platypush.message.event import Event
+from platypush.plugins.assistant import AssistantPlugin
+from platypush.utils import get_plugin_name_by_class
 
 
 class AssistantEvent(Event):
     """Base class for assistant events"""
 
-    def __init__(self, *args, assistant: Optional[str] = None, **kwargs):
+    def __init__(
+        self, *args, assistant: Optional[Union[str, AssistantPlugin]] = None, **kwargs
+    ):
         """
         :param assistant: Name of the assistant plugin that triggered the event.
         """
-        super().__init__(*args, assistant=assistant, **kwargs)
+        assistant = assistant or kwargs.get('assistant')
+        if assistant:
+            assistant = (
+                assistant
+                if isinstance(assistant, str)
+                else get_plugin_name_by_class(assistant.__class__)
+            )
+
+            kwargs['_assistant'] = assistant
+
+        super().__init__(*args, **kwargs)
 
     @property
-    def _assistant(self):
-        return (
-            get_plugin(self.args.get('assistant'))
-            if self.args.get('assistant')
-            else None
-        )
+    def assistant(self) -> Optional[AssistantPlugin]:
+        assistant = self.args.get('_assistant')
+        if not assistant:
+            return None
+
+        return get_plugin(assistant)
+
+    def as_dict(self):
+        evt_dict = super().as_dict()
+        evt_args = {**evt_dict['args']}
+        assistant = evt_args.pop('_assistant', None)
+        if assistant:
+            evt_args['assistant'] = assistant
+
+        return {
+            **evt_dict,
+            'args': evt_args,
+        }
 
 
 class ConversationStartEvent(AssistantEvent):
@@ -95,8 +121,8 @@ class SpeechRecognizedEvent(AssistantEvent):
         """
 
         result = super().matches_condition(condition)
-        if result.is_match and self._assistant and 'phrase' in condition.args:
-            self._assistant.stop_conversation()
+        if result.is_match and self.assistant and 'phrase' in condition.args:
+            self.assistant.stop_conversation()
 
         return result
 

platypush/plugins/__init__.py

@@ -122,6 +122,12 @@ class Plugin(EventGenerator, ExtensionWithManifest):  # lgtm [py/missing-call-to
         assert entities, 'entities plugin not initialized'
         return entities
 
+    def __str__(self):
+        """
+        :return: The qualified name of the plugin.
+        """
+        return get_plugin_name_by_class(self.__class__)
+
     def run(self, method, *args, **kwargs):
         assert (
             method in self.registered_actions

platypush/plugins/assistant/__init__.py

@@ -8,24 +8,7 @@ from typing import Any, Collection, Dict, Optional, Type
 from platypush.context import get_bus, get_plugin
 from platypush.entities.assistants import Assistant
 from platypush.entities.managers.assistants import AssistantEntityManager
-from platypush.message.event.assistant import (
-    AlarmEndEvent,
-    AlarmStartedEvent,
-    AlertEndEvent,
-    AlertStartedEvent,
-    AssistantEvent,
-    ConversationEndEvent,
-    ConversationStartEvent,
-    ConversationTimeoutEvent,
-    HotwordDetectedEvent,
-    MicMutedEvent,
-    MicUnmutedEvent,
-    NoResponseEvent,
-    ResponseEvent,
-    SpeechRecognizedEvent,
-    TimerEndEvent,
-    TimerStartedEvent,
-)
+from platypush.message.event import Event as AppEvent
 from platypush.plugins import Plugin, action
 from platypush.utils import get_plugin_name_by_class
 
@@ -182,6 +165,17 @@ class AssistantPlugin(Plugin, AssistantEntityManager, ABC):
         self.publish_entities([self])
         return asdict(self._state)
 
+    @action
+    def render_response(self, text: str, *_, **__):
+        """
+        Render a response text as audio over the configured TTS plugin.
+
+        :param text: Text to render.
+        """
+        self._on_response_render_start(text)
+        self._render_response(text)
+        self._on_response_render_end()
+
     def _get_tts_plugin(self):
         if not self.tts_plugin:
             return None
@@ -201,11 +195,13 @@ class AssistantPlugin(Plugin, AssistantEntityManager, ABC):
 
         audio.play(self._conversation_start_sound)
 
-    def _send_event(self, event_type: Type[AssistantEvent], **kwargs):
+    def _send_event(self, event_type: Type[AppEvent], **kwargs):
         self.publish_entities([self])
         get_bus().post(event_type(assistant=self._plugin_name, **kwargs))
 
     def _on_conversation_start(self):
+        from platypush.message.event.assistant import ConversationStartEvent
+
         self._last_response = None
         self._last_query = None
         self._conversation_running.set()
@@ -213,66 +209,98 @@ class AssistantPlugin(Plugin, AssistantEntityManager, ABC):
         self._play_conversation_start_sound()
 
     def _on_conversation_end(self):
+        from platypush.message.event.assistant import ConversationEndEvent
+
         self._conversation_running.clear()
         self._send_event(ConversationEndEvent)
 
     def _on_conversation_timeout(self):
+        from platypush.message.event.assistant import ConversationTimeoutEvent
+
         self._last_response = None
         self._last_query = None
         self._conversation_running.clear()
         self._send_event(ConversationTimeoutEvent)
 
     def _on_no_response(self):
+        from platypush.message.event.assistant import NoResponseEvent
+
         self._last_response = None
         self._conversation_running.clear()
         self._send_event(NoResponseEvent)
 
-    def _on_reponse_rendered(self, text: Optional[str]):
+    def _on_response_render_start(self, text: Optional[str]):
+        from platypush.message.event.assistant import ResponseEvent
+
         self._last_response = text
         self._send_event(ResponseEvent, response_text=text)
-        tts = self._get_tts_plugin()
 
+    def _render_response(self, text: Optional[str]):
+        tts = self._get_tts_plugin()
         if tts and text:
             self.stop_conversation()
             tts.say(text=text, **self.tts_plugin_args)
 
+    def _on_response_render_end(self):
+        pass
+
     def _on_hotword_detected(self, hotword: Optional[str]):
+        from platypush.message.event.assistant import HotwordDetectedEvent
+
         self._send_event(HotwordDetectedEvent, hotword=hotword)
 
     def _on_speech_recognized(self, phrase: Optional[str]):
+        from platypush.message.event.assistant import SpeechRecognizedEvent
+
         phrase = (phrase or '').lower().strip()
         self._last_query = phrase
         self._send_event(SpeechRecognizedEvent, phrase=phrase)
 
     def _on_alarm_start(self):
+        from platypush.message.event.assistant import AlarmStartedEvent
+
         self._cur_alert_type = AlertType.ALARM
         self._send_event(AlarmStartedEvent)
 
     def _on_alarm_end(self):
+        from platypush.message.event.assistant import AlarmEndEvent
+
         self._cur_alert_type = None
         self._send_event(AlarmEndEvent)
 
     def _on_timer_start(self):
+        from platypush.message.event.assistant import TimerStartedEvent
+
         self._cur_alert_type = AlertType.TIMER
         self._send_event(TimerStartedEvent)
 
     def _on_timer_end(self):
+        from platypush.message.event.assistant import TimerEndEvent
+
         self._cur_alert_type = None
         self._send_event(TimerEndEvent)
 
     def _on_alert_start(self):
+        from platypush.message.event.assistant import AlertStartedEvent
+
         self._cur_alert_type = AlertType.ALERT
         self._send_event(AlertStartedEvent)
 
     def _on_alert_end(self):
+        from platypush.message.event.assistant import AlertEndEvent
+
         self._cur_alert_type = None
         self._send_event(AlertEndEvent)
 
     def _on_mute(self):
+        from platypush.message.event.assistant import MicMutedEvent
+
         self._is_muted = True
         self._send_event(MicMutedEvent)
 
     def _on_unmute(self):
+        from platypush.message.event.assistant import MicUnmutedEvent
+
         self._is_muted = False
         self._send_event(MicUnmutedEvent)
 

platypush/plugins/assistant/picovoice/__init__.py

@@ -1,7 +1,10 @@
+import os
 from typing import Optional, Sequence
 
+from platypush.context import get_plugin
 from platypush.plugins import RunnablePlugin, action
 from platypush.plugins.assistant import AssistantPlugin
+from platypush.plugins.tts.picovoice import TtsPicovoicePlugin
 
 from ._assistant import Assistant
 from ._state import AssistantState
@@ -96,7 +99,12 @@ class AssistantPicovoicePlugin(AssistantPlugin, RunnablePlugin):
             using a language other than English, you can provide the path to the
             model file for that language. Model files are available for all the
             supported languages through the `Picovoice repository
-            <https://github.com/Picovoice/porcupine/tree/master/lib/common>`_.
+            <https://github.com/Picovoice/cheetah/tree/master/lib/common>`_.
+            You can also use the `Picovoice console
+            <https://console.picovoice.ai/cat>`_
+            to train your custom models. You can use a base model and fine-tune
+            it by boosting the detection of your own words and phrases and edit
+            the phonetic representation of the words you want to detect.
         :param endpoint_duration: If set, the assistant will stop listening when
             no speech is detected for the specified duration (in seconds) after
             the end of an utterance.
@@ -146,15 +154,43 @@ class AssistantPicovoicePlugin(AssistantPlugin, RunnablePlugin):
             'on_hotword_detected': self._on_hotword_detected,
         }
 
+    @property
+    def tts(self) -> TtsPicovoicePlugin:
+        p = get_plugin('tts.picovoice')
+        assert p, 'Picovoice TTS plugin not configured/found'
+        return p
+
+    def _get_tts_plugin(self) -> TtsPicovoicePlugin:
+        return self.tts
+
+    def _on_response_render_start(self, text: Optional[str]):
+        if self._assistant:
+            self._assistant.set_responding(True)
+        return super()._on_response_render_start(text)
+
+    def _on_response_render_end(self):
+        if self._assistant:
+            self._assistant.set_responding(False)
+
+        return super()._on_response_render_end()
+
     @action
-    def start_conversation(self, *_, **__):
+    def start_conversation(self, *_, model_file: Optional[str] = None, **__):
         """
-        Programmatically start a conversation with the assistant
+        Programmatically start a conversation with the assistant.
+
+        :param model_file: Override the model file to be used to detect speech
+            in this conversation. If not set, the configured
+            ``speech_model_path`` will be used.
         """
         if not self._assistant:
             self.logger.warning('Assistant not initialized')
             return
 
+        if model_file:
+            model_file = os.path.expanduser(model_file)
+
+        self._assistant.override_speech_model(model_file)
         self._assistant.state = AssistantState.DETECTING_SPEECH
 
     @action
@@ -166,6 +202,8 @@ class AssistantPicovoicePlugin(AssistantPlugin, RunnablePlugin):
             self.logger.warning('Assistant not initialized')
             return
 
+        self._assistant.override_speech_model(None)
+
         if self._assistant.hotword_enabled:
             self._assistant.state = AssistantState.DETECTING_HOTWORD
         else:
@@ -215,7 +253,8 @@ class AssistantPicovoicePlugin(AssistantPlugin, RunnablePlugin):
             with Assistant(**self._assistant_args) as self._assistant:
                 try:
                     for event in self._assistant:
-                        self.logger.debug('Picovoice assistant event: %s', event)
+                        if event is not None:
+                            self.logger.debug('Picovoice assistant event: %s', event)
                 except KeyboardInterrupt:
                     break
                 except Exception as e:

platypush/plugins/assistant/picovoice/_assistant.py

@@ -9,11 +9,13 @@ import pvleopard
 import pvporcupine
 import pvrhino
 
+from platypush.context import get_plugin
 from platypush.message.event.assistant import (
     ConversationTimeoutEvent,
     HotwordDetectedEvent,
     SpeechRecognizedEvent,
 )
+from platypush.plugins.tts.picovoice import TtsPicovoicePlugin
 
 from ._context import ConversationContext
 from ._recorder import AudioRecorder
@@ -25,6 +27,7 @@ class Assistant:
     A facade class that wraps the Picovoice engines under an assistant API.
     """
 
+    @staticmethod
     def _default_callback(*_, **__):
         pass
 
@@ -60,12 +63,14 @@ class Assistant:
         self.keywords = list(keywords or [])
         self.keyword_paths = None
         self.keyword_model_path = None
+        self._responding = Event()
         self.frame_expiration = frame_expiration
-        self.speech_model_path = speech_model_path
         self.endpoint_duration = endpoint_duration
         self.enable_automatic_punctuation = enable_automatic_punctuation
         self.start_conversation_on_hotword = start_conversation_on_hotword
         self.audio_queue_size = audio_queue_size
+        self._speech_model_path = speech_model_path
+        self._speech_model_path_override = None
 
         self._on_conversation_start = on_conversation_start
         self._on_conversation_end = on_conversation_end
@@ -103,11 +108,32 @@ class Assistant:
 
                 self.keyword_model_path = keyword_model_path
 
-        self._cheetah: Optional[pvcheetah.Cheetah] = None
+        # Model path -> model instance cache
+        self._cheetah = {}
         self._leopard: Optional[pvleopard.Leopard] = None
         self._porcupine: Optional[pvporcupine.Porcupine] = None
         self._rhino: Optional[pvrhino.Rhino] = None
 
+    @property
+    def is_responding(self):
+        return self._responding.is_set()
+
+    @property
+    def speech_model_path(self):
+        return self._speech_model_path_override or self._speech_model_path
+
+    @property
+    def tts(self) -> TtsPicovoicePlugin:
+        p = get_plugin('tts.picovoice')
+        assert p, 'Picovoice TTS plugin not configured/found'
+        return p
+
+    def set_responding(self, responding: bool):
+        if responding:
+            self._responding.set()
+        else:
+            self._responding.clear()
+
     def should_stop(self):
         return self._stop_event.is_set()
 
@@ -130,12 +156,18 @@ class Assistant:
             return
 
         if prev_state == AssistantState.DETECTING_SPEECH:
+            self.tts.stop()
             self._ctx.stop()
+            self._speech_model_path_override = None
             self._on_conversation_end()
         elif new_state == AssistantState.DETECTING_SPEECH:
             self._ctx.start()
             self._on_conversation_start()
 
+        if new_state == AssistantState.DETECTING_HOTWORD:
+            self.tts.stop()
+            self._ctx.reset()
+
     @property
     def porcupine(self) -> Optional[pvporcupine.Porcupine]:
         if not self.hotword_enabled:
@@ -159,7 +191,7 @@ class Assistant:
         if not self.stt_enabled:
             return None
 
-        if not self._cheetah:
+        if not self._cheetah.get(self.speech_model_path):
             args: Dict[str, Any] = {'access_key': self._access_key}
             if self.speech_model_path:
                 args['model_path'] = self.speech_model_path
@@ -168,20 +200,22 @@ class Assistant:
             if self.enable_automatic_punctuation:
                 args['enable_automatic_punctuation'] = self.enable_automatic_punctuation
 
-            self._cheetah = pvcheetah.create(**args)
+            self._cheetah[self.speech_model_path] = pvcheetah.create(**args)
 
-        return self._cheetah
+        return self._cheetah[self.speech_model_path]
 
     def __enter__(self):
+        """
+        Get the assistant ready to start processing audio frames.
+        """
         if self.should_stop():
             return self
 
         if self._recorder:
             self.logger.info('A recording stream already exists')
-        elif self.porcupine or self.cheetah:
+        elif self.hotword_enabled or self.stt_enabled:
             sample_rate = (self.porcupine or self.cheetah).sample_rate  # type: ignore
             frame_length = (self.porcupine or self.cheetah).frame_length  # type: ignore
-
             self._recorder = AudioRecorder(
                 stop_event=self._stop_event,
                 sample_rate=sample_rate,
@@ -190,6 +224,9 @@ class Assistant:
                 channels=1,
             )
 
+            if self.stt_enabled:
+                self._cheetah[self.speech_model_path] = self.cheetah
+
             self._recorder.__enter__()
 
             if self.porcupine:
@@ -200,15 +237,18 @@ class Assistant:
         return self
 
     def __exit__(self, *_):
+        """
+        Stop the assistant and release all resources.
+        """
         if self._recorder:
             self._recorder.__exit__(*_)
             self._recorder = None
 
         self.state = AssistantState.IDLE
-
-        if self._cheetah:
-            self._cheetah.delete()
-            self._cheetah = None
+        for model in [*self._cheetah.keys()]:
+            cheetah = self._cheetah.pop(model, None)
+            if cheetah:
+                cheetah.delete()
 
         if self._leopard:
             self._leopard.delete()
@@ -223,9 +263,15 @@ class Assistant:
             self._rhino = None
 
     def __iter__(self):
+        """
+        Iterate over processed assistant events.
+        """
         return self
 
     def __next__(self):
+        """
+        Process the next audio frame and return the corresponding event.
+        """
         has_data = False
         if self.should_stop() or not self._recorder:
             raise StopIteration
@@ -242,10 +288,10 @@ class Assistant:
                 )
                 continue  # The audio frame is too old
 
-            if self.porcupine and self.state == AssistantState.DETECTING_HOTWORD:
+            if self.hotword_enabled and self.state == AssistantState.DETECTING_HOTWORD:
                 return self._process_hotword(frame)
 
-            if self.cheetah and self.state == AssistantState.DETECTING_SPEECH:
+            if self.stt_enabled and self.state == AssistantState.DETECTING_SPEECH:
                 return self._process_speech(frame)
 
         raise StopIteration
@@ -262,6 +308,7 @@ class Assistant:
             if self.start_conversation_on_hotword:
                 self.state = AssistantState.DETECTING_SPEECH
 
+            self.tts.stop()
             self._on_hotword_detected(hotword=self.keywords[keyword_index])
             return HotwordDetectedEvent(hotword=self.keywords[keyword_index])
 
@@ -275,23 +322,20 @@ class Assistant:
         partial_transcript, self._ctx.is_final = self.cheetah.process(frame)
 
         if partial_transcript:
-            self._ctx.partial_transcript += partial_transcript
+            self._ctx.transcript += partial_transcript
             self.logger.info(
                 'Partial transcript: %s, is_final: %s',
-                self._ctx.partial_transcript,
+                self._ctx.transcript,
                 self._ctx.is_final,
             )
 
         if self._ctx.is_final or self._ctx.timed_out:
-            phrase = ''
-            if self.cheetah:
-                phrase = self.cheetah.flush()
-
-            self._ctx.partial_transcript += phrase
-            phrase = self._ctx.partial_transcript
+            phrase = self.cheetah.flush() or ''
+            self._ctx.transcript += phrase
+            phrase = self._ctx.transcript
             phrase = phrase[:1].lower() + phrase[1:]
 
-            if self._ctx.is_final or phrase:
+            if phrase:
                 event = SpeechRecognizedEvent(phrase=phrase)
                 self._on_speech_recognized(phrase=phrase)
             else:
@@ -304,5 +348,8 @@ class Assistant:
 
         return event
 
+    def override_speech_model(self, model_path: Optional[str]):
+        self._speech_model_path_override = model_path
+
 
 # vim:sw=4:ts=4:et:

platypush/plugins/assistant/picovoice/_context.py

@@ -9,7 +9,7 @@ class ConversationContext:
     Context of the conversation process.
     """
 
-    partial_transcript: str = ''
+    transcript: str = ''
     is_final: bool = False
     timeout: Optional[float] = None
     t_start: Optional[float] = None
@@ -24,7 +24,7 @@ class ConversationContext:
         self.t_end = time()
 
     def reset(self):
-        self.partial_transcript = ''
+        self.transcript = ''
         self.is_final = False
         self.t_start = None
         self.t_end = None
@@ -32,11 +32,17 @@ class ConversationContext:
     @property
     def timed_out(self):
         return (
-            not self.partial_transcript
+            not self.transcript
             and not self.is_final
             and self.timeout
             and self.t_start
             and time() - self.t_start > self.timeout
+        ) or (
+            self.transcript
+            and not self.is_final
+            and self.timeout
+            and self.t_start
+            and time() - self.t_start > self.timeout * 2
         )
 
 

platypush/plugins/assistant/picovoice/manifest.yaml

@@ -12,7 +12,14 @@ manifest:
     - platypush.message.event.assistant.ResponseEvent
     - platypush.message.event.assistant.SpeechRecognizedEvent
   install:
+    apk:
+      - ffmpeg
+    apt:
+      - ffmpeg
+    dnf:
+      - ffmpeg
     pacman:
+      - ffmpeg
       - python-sounddevice
     pip:
       - pvcheetah

platypush/plugins/tts/picovoice/__init__.py

@@ -0,0 +1,138 @@
+import os
+from threading import RLock
+from typing import Optional
+
+import numpy as np
+import pvorca
+import sounddevice as sd
+
+from platypush.config import Config
+from platypush.plugins import action
+from platypush.plugins.tts import TtsPlugin
+
+
+class TtsPicovoicePlugin(TtsPlugin):
+    """
+    This TTS plugin enables you to render text as audio using `Picovoice
+    <https://picovoice.ai>`_'s (still experimental) `Orca TTS engine
+    <https://github.com/Picovoice/orca>`_.
+
+    Take a look at
+    :class:`platypush.plugins.assistant.picovoice.AssistantPicovoicePlugin`
+    for details on how to sign up for a Picovoice account and get the API key.
+
+    Also note that using the TTS features requires you to select Orca from the
+    list of products available for your account on the `Picovoice console
+    <https://console.picovoice.ai>`_.
+    """
+
+    def __init__(
+        self,
+        access_key: Optional[str] = None,
+        model_path: Optional[str] = None,
+        **kwargs,
+    ):
+        """
+        :param access_key: Picovoice access key. If it's not specified here,
+            then it must be specified on the configuration of
+            :class:`platypush.plugins.assistant.picovoice.AssistantPicovoicePlugin`.
+        :param model_path: Path of the TTS model file (default: use the default
+            English model).
+        """
+        super().__init__(**kwargs)
+        if not access_key:
+            access_key = Config.get('assistant.picovoice', {}).get('access_key')
+            assert (
+                access_key
+            ), 'No access key specified and no assistant.picovoice plugin found'
+
+        self.model_path = model_path
+        self.access_key = access_key
+        if model_path:
+            model_path = os.path.expanduser(model_path)
+
+        self._stream: Optional[sd.OutputStream] = None
+        self._stream_lock = RLock()
+
+    def _play_audio(self, orca: pvorca.Orca, pcm: np.ndarray):
+        with self._stream_lock:
+            self.stop()
+            self._stream = sd.OutputStream(
+                samplerate=orca.sample_rate,
+                channels=1,
+                dtype='int16',
+            )
+
+        try:
+            self._stream.start()
+            self._stream.write(pcm)
+        except Exception as e:
+            self.logger.warning('Error playing audio: %s: %s', type(e), str(e))
+        finally:
+            try:
+                self.stop()
+                self._stream.close()
+            except Exception as e:
+                self.logger.warning(
+                    'Error stopping audio stream: %s: %s', type(e), str(e)
+                )
+            finally:
+                if self._stream:
+                    self._stream = None
+
+    def get_orca(self, model_path: Optional[str] = None):
+        if not model_path:
+            model_path = self.model_path
+        if model_path:
+            model_path = os.path.expanduser(model_path)
+
+        return pvorca.create(access_key=self.access_key, model_path=model_path)
+
+    @action
+    def say(
+        self,
+        text: str,
+        *_,
+        output_file: Optional[str] = None,
+        speech_rate: Optional[float] = None,
+        model_path: Optional[str] = None,
+        **__,
+    ):
+        """
+        Say some text.
+
+        :param text: Text to say.
+        :param output_file: If set, save the audio to the specified file.
+            Otherwise play it.
+        :param speech_rate: Speech rate (default: None).
+        :param model_path: Path of the TTS model file (default: use the default
+            configured model).
+        """
+        orca = self.get_orca(model_path=model_path)
+        if output_file:
+            orca.synthesize_to_file(
+                text, os.path.expanduser(output_file), speech_rate=speech_rate
+            )
+            return
+
+        self._play_audio(
+            orca=orca,
+            pcm=np.array(
+                orca.synthesize(text, speech_rate=speech_rate),
+                dtype='int16',
+            ),
+        )
+
+    @action
+    def stop(self):
+        """
+        Stop the currently playing audio.
+        """
+        with self._stream_lock:
+            if not self._stream:
+                return
+
+            self._stream.stop()
+
+
+# vim:sw=4:ts=4:et:

platypush/plugins/tts/picovoice/manifest.yaml

@@ -0,0 +1,22 @@
+manifest:
+  events: {}
+  install:
+    apk:
+      - ffmpeg
+      - py3-numpy
+    apt:
+      - ffmpeg
+      - python3-numpy
+    dnf:
+      - ffmpeg
+      - python-numpy
+    pacman:
+      - ffmpeg
+      - python-numpy
+      - python-sounddevice
+    pip:
+      - numpy
+      - pvorca
+      - sounddevice
+  package: platypush.plugins.tts.picovoice
+  type: plugin