livekit-plugins-aws 1.1.4__py3-none-any.whl → 1.1.6__py3-none-any.whl

livekit/plugins/aws/experimental/realtime/turn_tracker.py

@@ -0,0 +1,163 @@
+from __future__ import annotations
+
+import datetime
+import enum
+import uuid
+from dataclasses import dataclass, field
+from typing import Any, Callable
+
+from livekit.agents import llm
+
+from ...log import logger
+
+
+class _Phase(enum.Enum):
+    IDLE = 0  # waiting for the USER to begin speaking
+    USER_SPEAKING = 1  # still receiving USER text+audio blocks
+    USER_FINISHED = 2  # first ASSISTANT speculative block observed
+    ASSISTANT_RESPONDING = 3  # ASSISTANT audio/text streaming
+    DONE = 4  # assistant audio ended (END_TURN) or barge-in (INTERRUPTED)
+
+
+# note: b/c user ASR text is transcribed server-side, a single turn constitutes
+# both the user and agent's speech
+@dataclass
+class _Turn:
+    turn_id: int
+    input_id: str = field(default_factory=lambda: str(uuid.uuid4()))
+    created: datetime.datetime = field(default_factory=datetime.datetime.utcnow)
+    transcript: list[str] = field(default_factory=list)
+
+    phase: _Phase = _Phase.IDLE
+    ev_input_started: bool = False
+    ev_input_stopped: bool = False
+    ev_trans_completed: bool = False
+    ev_generation_sent: bool = False
+
+    def add_partial_text(self, text: str) -> None:
+        self.transcript.append(text)
+
+    @property
+    def curr_transcript(self) -> str:
+        return " ".join(self.transcript)
+
+
+class _TurnTracker:
+    def __init__(
+        self,
+        emit_fn: Callable[[str, Any], None],
+        emit_generation_fn: Callable[[], None],
+    ):
+        self._emit = emit_fn
+        self._turn_idx = 0
+        self._curr_turn: _Turn | None = None
+        self._emit_generation_fn = emit_generation_fn
+
+    # --------------------------------------------------------
+    #  PUBLIC ENTRY POINT
+    # --------------------------------------------------------
+    def feed(self, event: dict) -> None:
+        turn = self._ensure_turn()
+        kind = _classify(event)
+
+        if kind == "USER_TEXT_PARTIAL":
+            turn.add_partial_text(event["event"]["textOutput"]["content"])
+            self._maybe_emit_input_started(turn)
+            self._emit_transcript_updated(turn)
+            # note: cannot invoke self._maybe_input_stopped() here
+            # b/c there is no way to know if the user is done speaking
+
+        # will always be correlated b/c generate_reply() is a stub
+        # user ASR text ends when agent's ASR speculative text begins
+        # corresponds to beginning of agent's turn
+        elif kind == "TOOL_OUTPUT_CONTENT_START" or kind == "ASSISTANT_SPEC_START":
+            # must be a maybe methods b/c agent can chain multiple tool calls
+            self._maybe_emit_input_stopped(turn)
+            self._maybe_emit_transcript_completed(turn)
+            self._maybe_emit_generation_created(turn)
+
+        elif kind == "BARGE_IN":
+            logger.debug(f"BARGE-IN DETECTED IN TURN TRACKER: {turn}")
+            # start new turn immediately to make interruptions snappier
+            self._emit("input_speech_started", llm.InputSpeechStartedEvent())
+            turn.phase = _Phase.DONE
+
+        elif kind == "ASSISTANT_AUDIO_END":
+            if event["event"]["contentEnd"]["stopReason"] == "END_TURN":
+                turn.phase = _Phase.DONE
+
+        if turn.phase is _Phase.DONE:
+            self._curr_turn = None
+
+    def _ensure_turn(self) -> _Turn:
+        if self._curr_turn is None:
+            self._turn_idx += 1
+            self._curr_turn = _Turn(turn_id=self._turn_idx)
+        return self._curr_turn
+
+    def _maybe_emit_input_started(self, turn: _Turn) -> None:
+        if not turn.ev_input_started:
+            turn.ev_input_started = True
+            self._emit("input_speech_started", llm.InputSpeechStartedEvent())
+            turn.phase = _Phase.USER_SPEAKING
+
+    def _maybe_emit_input_stopped(self, turn: _Turn) -> None:
+        if not turn.ev_input_stopped:
+            turn.ev_input_stopped = True
+            self._emit(
+                "input_speech_stopped", llm.InputSpeechStoppedEvent(user_transcription_enabled=True)
+            )
+            turn.phase = _Phase.USER_FINISHED
+
+    def _emit_transcript_updated(self, turn: _Turn) -> None:
+        self._emit(
+            "input_audio_transcription_completed",
+            llm.InputTranscriptionCompleted(
+                item_id=turn.input_id,
+                transcript=turn.curr_transcript,
+                is_final=False,
+            ),
+        )
+
+    def _maybe_emit_transcript_completed(self, turn: _Turn) -> None:
+        if not turn.ev_trans_completed:
+            turn.ev_trans_completed = True
+            self._emit(
+                "input_audio_transcription_completed",
+                # Q: does input_id need to match /w the _ResponseGeneration.input_id?
+                llm.InputTranscriptionCompleted(
+                    item_id=turn.input_id,
+                    transcript=turn.curr_transcript,
+                    is_final=True,
+                ),
+            )
+
+    def _maybe_emit_generation_created(self, turn: _Turn) -> None:
+        if not turn.ev_generation_sent:
+            turn.ev_generation_sent = True
+            self._emit_generation_fn()
+            turn.phase = _Phase.ASSISTANT_RESPONDING
+
+
+def _classify(ev: dict) -> str:
+    e = ev.get("event", {})
+    if "textOutput" in e and e["textOutput"]["role"] == "USER":
+        return "USER_TEXT_PARTIAL"
+
+    if "contentStart" in e and e["contentStart"]["type"] == "TOOL":
+        return "TOOL_OUTPUT_CONTENT_START"
+
+    if "contentStart" in e and e["contentStart"]["role"] == "ASSISTANT":
+        add = e["contentStart"].get("additionalModelFields", "")
+        if "SPECULATIVE" in add:
+            return "ASSISTANT_SPEC_START"
+
+    if "textOutput" in e and e["textOutput"]["content"] == '{ "interrupted" : true }':
+        return "BARGE_IN"
+
+    # note: there cannot be any audio events for the user in the output event loop
+    # therefore, we know that the audio event must be for the assistant
+    if "contentEnd" in e and e["contentEnd"]["type"] == "AUDIO":
+        return "ASSISTANT_AUDIO_END"
+
+    return ""

livekit/plugins/aws/log.py

@@ -1,3 +1,7 @@
 import logging
 
 logger = logging.getLogger("livekit.plugins.aws")
+smithy_logger = logging.getLogger("smithy_aws_event_stream.aio")
+smithy_logger.setLevel(logging.INFO)
+bedrock_client_logger = logging.getLogger("aws_sdk_bedrock_runtime.client")
+bedrock_client_logger.setLevel(logging.INFO)

livekit/plugins/aws/version.py

@@ -12,4 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "1.1.4"
+__version__ = "1.1.6"

{livekit_plugins_aws-1.1.4.dist-info → livekit_plugins_aws-1.1.6.dist-info}/METADATA

@@ -1,13 +1,13 @@
 Metadata-Version: 2.4
 Name: livekit-plugins-aws
-Version: 1.1.4
+Version: 1.1.6
 Summary: LiveKit Agents Plugin for services from AWS
 Project-URL: Documentation, https://docs.livekit.io
 Project-URL: Website, https://livekit.io/
 Project-URL: Source, https://github.com/livekit/agents
 Author-email: LiveKit <hello@livekit.io>
 License-Expression: Apache-2.0
-Keywords: audio,aws,livekit,realtime,video,webrtc
+Keywords: audio,aws,livekit,nova,realtime,sonic,video,webrtc
 Classifier: Intended Audience :: Developers
 Classifier: License :: OSI Approved :: Apache Software License
 Classifier: Programming Language :: Python :: 3
@@ -20,12 +20,15 @@ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.9.0
 Requires-Dist: aioboto3>=14.1.0
 Requires-Dist: amazon-transcribe>=0.6.2
-Requires-Dist: livekit-agents>=1.1.4
+Requires-Dist: livekit-agents>=1.1.6
+Provides-Extra: realtime
+Requires-Dist: aws-sdk-bedrock-runtime==0.0.2; (python_version >= '3.12') and extra == 'realtime'
+Requires-Dist: boto3>1.35.10; extra == 'realtime'
 Description-Content-Type: text/markdown
 
 # AWS plugin for LiveKit Agents
 
-Support for AWS AI including Bedrock, Polly, and Transcribe.
+Support for AWS AI including Bedrock, Polly, Transcribe and optionally Nova Sonic (realtime STS model).
 
 See [https://docs.livekit.io/agents/integrations/aws/](https://docs.livekit.io/agents/integrations/aws/) for more information.
 
@@ -33,8 +36,15 @@ See [https://docs.livekit.io/agents/integrations/aws/](https://docs.livekit.io/a
 
 ```bash
 pip install livekit-plugins-aws
+
+# for access to Nova Sonic
+pip install livekit-plugins-aws[realtime]
 ```
 
 ## Pre-requisites
 
-You'll need to specify an AWS Access Key and a Deployment Region. They can be set as environment variables: `AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY` and `AWS_DEFAULT_REGION`, respectively.
+You'll need to specify an AWS Access Key and a Deployment Region. They can be set as environment variables: `AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY` and `AWS_DEFAULT_REGION`, respectively.
+
+## Example
+
+For an example of the realtime STS model, Nova Sonic, see: https://github.com/livekit/agents/blob/main/examples/voice_agents/realtime_joke_teller.py

livekit_plugins_aws-1.1.6.dist-info/RECORD

@@ -0,0 +1,17 @@
+livekit/plugins/aws/__init__.py,sha256=dCZISj1yZG0WZTojk3sU-Ub4PK1ThCVhamrl9k_NbBw,2047
+livekit/plugins/aws/llm.py,sha256=pSbO7SaqYZYJ-3JGOmyMTmTLLcjoyrZy-j5BSkFOHU8,11922
+livekit/plugins/aws/log.py,sha256=S5ICcsnwshZhMG0HPmc_lI3mtHmcY4oQMJBsnnho-bM,289
+livekit/plugins/aws/models.py,sha256=ztI9n3r8OHGlQewWXiyQP3SaavFKHkw-VVuK9Xz4K1I,696
+livekit/plugins/aws/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+livekit/plugins/aws/stt.py,sha256=PSR89aN28wm4i83yEdhkDJ9xzM0CsNIKrc3v3EbPndQ,9018
+livekit/plugins/aws/tts.py,sha256=T5dVpTuIuzQimYNnkfXi5dRLmRldWySL4IcbkXjmJLM,6083
+livekit/plugins/aws/utils.py,sha256=nA5Ua1f4T-25Loar6EvlrKTXI9N-zpTIH7cdQkwGyGI,1518
+livekit/plugins/aws/version.py,sha256=-bNd31cMcYCdhZCIKJ1-jtY4NgZvppVgKyzXAIzQtqM,600
+livekit/plugins/aws/experimental/realtime/__init__.py,sha256=mm_TGZc9QAWSO-VOO3PdE8Y5R6xlWckXRZuiFUIHa-Q,287
+livekit/plugins/aws/experimental/realtime/events.py,sha256=-pJrwVrH5AZFxa1eDbX5nDdnJMz4BNucNZlYUYLsP-Y,15853
+livekit/plugins/aws/experimental/realtime/pretty_printer.py,sha256=KN7KPrfQu8cU7ff34vFAtfrd1umUSTVNKXQU7D8AMiM,1442
+livekit/plugins/aws/experimental/realtime/realtime_model.py,sha256=em_3Fbp1qefF7cIIHc6ib1FLdD1MOGes2Lwq61o2wlk,59464
+livekit/plugins/aws/experimental/realtime/turn_tracker.py,sha256=bcufaap-coeIYuK3ct1Is9W_UoefGYRmnJu7Mn5DCYU,6002
+livekit_plugins_aws-1.1.6.dist-info/METADATA,sha256=ST8uYsoqQgHUVRCLC3BdkYdwALh3joYGRjblVKQgDrE,1989
+livekit_plugins_aws-1.1.6.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+livekit_plugins_aws-1.1.6.dist-info/RECORD,,

livekit_plugins_aws-1.1.4.dist-info/RECORD

@@ -1,12 +0,0 @@
-livekit/plugins/aws/__init__.py,sha256=fkbgTfNZc6z4VxbYGNdY73EoGvNuRcJiuD-OlUHvjHU,1322
-livekit/plugins/aws/llm.py,sha256=pSbO7SaqYZYJ-3JGOmyMTmTLLcjoyrZy-j5BSkFOHU8,11922
-livekit/plugins/aws/log.py,sha256=jFief0Xhv0n_F6sp6UFu9VKxs2bXNVGAfYGmEYfR_2Q,66
-livekit/plugins/aws/models.py,sha256=ztI9n3r8OHGlQewWXiyQP3SaavFKHkw-VVuK9Xz4K1I,696
-livekit/plugins/aws/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-livekit/plugins/aws/stt.py,sha256=PSR89aN28wm4i83yEdhkDJ9xzM0CsNIKrc3v3EbPndQ,9018
-livekit/plugins/aws/tts.py,sha256=T5dVpTuIuzQimYNnkfXi5dRLmRldWySL4IcbkXjmJLM,6083
-livekit/plugins/aws/utils.py,sha256=nA5Ua1f4T-25Loar6EvlrKTXI9N-zpTIH7cdQkwGyGI,1518
-livekit/plugins/aws/version.py,sha256=B5kiBdHIDDcwA9hDZC77YLcp8cT8rcpNKzBQ-kJeiYY,600
-livekit_plugins_aws-1.1.4.dist-info/METADATA,sha256=67ZScP6Cqs8dTwuQgvFn0_4BHT1wJjkKl-AOkDbIz5A,1529
-livekit_plugins_aws-1.1.4.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-livekit_plugins_aws-1.1.4.dist-info/RECORD,,