Merge pull request #242 from aws-samples/fix/python-sdk-backend

fix(python-backend): Fix Nova Sonic protocol and SDK v0.1.0 compatibi…

Author: krokoko

samples/speech-to-speech/backend/python_app/clients/bedrock_client.py

@@ -15,7 +15,6 @@
 import logging
 import warnings
 import os
-import time
 from aws_sdk_bedrock_runtime.client import (
     BedrockRuntimeClient,
     InvokeModelWithBidirectionalStreamOperationInput,
@@ -24,15 +23,10 @@
     InvokeModelWithBidirectionalStreamInputChunk,
     BidirectionalInputPayloadPart,
 )
-from aws_sdk_bedrock_runtime.config import (
-    Config,
-    HTTPAuthSchemeResolver,
-    SigV4AuthScheme,
-)
-from smithy_aws_core.credentials_resolvers.environment import (
-    EnvironmentCredentialsResolver,
-)
-from smithy_aws_core.credentials_resolvers.container import ContainerCredentialsResolver
+from aws_sdk_bedrock_runtime.config import Config
+from smithy_aws_core.auth.sigv4 import SigV4AuthScheme
+from smithy_aws_core.identity.environment import EnvironmentCredentialsResolver
+from smithy_aws_core.identity.container import ContainerCredentialsResolver
 from smithy_http.aio.aiohttp import AIOHTTPClient, AIOHTTPClientConfig
 
 # Configure logging
@@ -100,8 +94,7 @@ def initialize_client(self):
                 endpoint_uri=f"https://bedrock-runtime.{self.region}.amazonaws.com",
                 region=self.region,
                 aws_credentials_identity_resolver=resolver,
-                http_auth_scheme_resolver=HTTPAuthSchemeResolver(),
-                http_auth_schemes={"aws.auth#sigv4": SigV4AuthScheme()},
+                auth_schemes={"aws.auth#sigv4": SigV4AuthScheme(service="bedrock")},
             )
             self.bedrock_client = BedrockRuntimeClient(config=config)
             logger.info(

samples/speech-to-speech/backend/python_app/core/audio_processor.py

@@ -83,6 +83,40 @@ async def queue_audio(self, prompt_name: str, content_name: str, audio_base64: s
             )
             return
 
+        # Add audio validation to detect silence
+        try:
+            import base64
+            audio_bytes = base64.b64decode(audio_base64)
+            
+            # Check if audio is silence (all zeros or very low values)
+            if len(audio_bytes) > 0:
+                # Convert bytes to 16-bit signed integers for analysis
+                import struct
+                num_samples = len(audio_bytes) // 2  # 16-bit samples
+                
+                if num_samples > 0:
+                    # Unpack first few samples to check for silence
+                    sample_check_count = min(100, num_samples)
+                    samples = struct.unpack(f'{sample_check_count}h', audio_bytes[:sample_check_count * 2])
+                    
+                    # Calculate max absolute value
+                    max_val = max(abs(s) for s in samples)
+                    
+                    # Threshold for silence detection (16-bit audio range is -32768 to 32767)
+                    # Only warn on truly silent audio (complete zeros)
+                    if max_val == 0:
+                        self.logger.warning(
+                            "Detected completely silent audio chunk (all zeros). "
+                            "Frontend may not be sending real microphone audio."
+                        )
+                    elif max_val < 50:
+                        # Very low audio - only log at debug level
+                        self.logger.debug(f"Very low audio level - max amplitude: {max_val}")
+                    # Normal audio - no logging needed (reduces verbosity)
+                        
+        except Exception as e:
+            self.logger.error(f"Error validating audio: {e}")
+
         # Use put_nowait() like original code to avoid blocking/timing delays
         self.audio_input_queue.put_nowait(
             {

samples/speech-to-speech/backend/python_app/core/websocket_handler.py

@@ -123,12 +123,12 @@ async def handle_connection(self, websocket):
             # Start audio processor
             await self.audio_processor.start(self.bedrock_client, self.stream)
 
-            # Start response processing task
-            self.response_task = asyncio.create_task(self._process_responses(websocket))
-
             # Register message handlers
             self._register_message_handlers()
 
+            # Start response processing task AFTER setup is complete 
+            self.response_task = asyncio.create_task(self._process_responses(websocket))
+
             # Main message processing loop
             async for message in websocket:
                 await self._handle_message(websocket, message)
@@ -235,14 +235,42 @@ async def _handle_message(self, websocket, message):
                 list(data.get("event", {}).keys())[0] if "event" in data else None
             )
 
-            # Store session information
-            if event_type == "promptStart":
-                self.prompt_name = data["event"]["promptStart"]["promptName"]
-            elif (
-                event_type == "contentStart"
-                and data["event"]["contentStart"].get("type") == "AUDIO"
-            ):
-                self.audio_content_name = data["event"]["contentStart"]["contentName"]
+            # Event sequence tracking (reduced verbosity)
+            if event_type:
+                if event_type == "sessionStart":
+                    self.logger.info("Session started")
+                elif event_type == "promptStart":
+                    self.prompt_name = data["event"]["promptStart"]["promptName"]
+                    self.logger.debug("Prompt started")
+                elif event_type == "contentStart":
+                    content_type = data["event"]["contentStart"].get("type")
+                    role = data["event"]["contentStart"].get("role", "UNSPECIFIED")
+                    content_name = data["event"]["contentStart"].get("contentName")
+                    
+                    # Only log at debug level for routine content
+                    self.logger.debug(f"Content start: type={content_type}, role={role}")
+                    
+                    # Validate first content block has SYSTEM role
+                    if not hasattr(self, '_first_content_received'):
+                        self._first_content_received = True
+                        if role != "SYSTEM":
+                            self.logger.error(f"First content block must have SYSTEM role, received {role}")
+                            await websocket.send(json.dumps({
+                                "type": "error",
+                                "message": f"First content block must have SYSTEM role, received {role}"
+                            }))
+                    
+                    # Store audio content name
+                    if content_type == "AUDIO":
+                        self.audio_content_name = content_name
+                        if role != "USER":
+                            self.logger.warning(f"Audio content should have USER role, received {role}")
+                            
+                elif event_type == "sessionEnd":
+                    self.logger.info("Session ended")
+                # Other events logged at debug level only
+                elif event_type in ["textInput", "contentEnd", "promptEnd"]:
+                    self.logger.debug(f"Event: {event_type}")
 
             # Send event to Bedrock
             await self.bedrock_client.send_event(self.stream, data)

samples/speech-to-speech/backend/python_app/events/s2s_events.py

@@ -157,6 +157,7 @@ def content_start_audio(
                     "promptName": prompt_name,
                     "contentName": content_name,
                     "type": "AUDIO",
+                    "role": "USER",  # Required by Nova Sonic API
                     "interactive": True,
                     "audioInputConfiguration": audio_input_config,
                 }

samples/speech-to-speech/backend/python_app/requirements.txt

@@ -1,6 +1,6 @@
 # Core dependencies
-aws_sdk_bedrock_runtime==0.0.2
-smithy-aws-core==0.0.3
+aws_sdk_bedrock_runtime==0.1.0
+smithy-aws-core==0.1.0
 python-dotenv>=1.1.0
 aiohttp>=3.8.0