- Newest
- Most votes
- Most comments
The workaround is to send a silent (pre-recorded) audio clip to Sonic to initiate the conversation.
answered a year ago
To modify the behavior of Amazon Nova Sonic so that it initiates the conversation with a welcome address after the "Start Conversation" button is clicked, you'll need to adjust the event flow in your application. Here's how you can achieve this:
-
After the "Start Conversation" button is clicked, instead of waiting for user input, immediately send a system prompt event to Nova Sonic. This prompt should instruct the model to start with a welcome message.
-
In your system prompt, include instructions for the model to begin the conversation. For example:
"You are a friendly assistant. Begin the conversation by welcoming the user with a brief greeting. Then, ask how you can help them today."
-
After sending the system prompt, you should immediately send a content start event followed by an empty user input event. This simulates the user's turn in the conversation without actually requiring user input.
-
Nova Sonic will then respond with its welcome message, initiating the conversation.
-
After this initial exchange, the normal back-and-forth conversation can proceed as usual.
Remember that Nova Sonic uses an event-driven architecture with a bidirectional stream API. You'll need to ensure that you're sending the correct sequence of events to trigger this behavior:
- Send the sessionStart event
- Send the promptStart event
- Send the contentStart event for the system prompt
- Send the textInput event with your system instructions
- Send the contentEnd event for the system prompt
- Send a contentStart event for user input
- Immediately send a contentEnd event for user input (simulating an empty user turn)
By following this sequence, you can make Nova Sonic initiate the conversation with a welcome message as soon as the "Start Conversation" button is clicked, without waiting for the user to speak first.
Sources
Handling input events with the bidirectional API - Amazon Nova
System prompt authoring guidelines and examples - Amazon Nova
Using the Amazon Nova Sonic Speech-to-Speech model - Amazon Nova
Thanks Oleksii Bebych(https://repost.aws/community/users/USyYP5VPAQSSy-rWgsNiZjCw), really awesome, one question regrading this issue, can i changes the code in either python or js or both. If yes means, then which python i need to modify the order, it will be either "s2s_events.py" or "s2s_session_manager.py"?
Code link --> "https://github.com/aws-samples/amazon-nova-samples/tree/main/speech-to-speech/workshops"
The automatically generated suggestion doesn't work. The empty contentStart + contentEnd suggested in steps 6 and 7 causes the error: "Cannot end content [UUID] as no content data was received".
Does anybody know how to make the Nova Sonic assistant speak first at the start of a conversation, rather than waiting for the user to speak first?
I spent a day trying every possible way to get this working. It doesn’t work.
Relevant content
asked a year ago
asked 8 months ago
asked 10 months ago
- AWS OFFICIALUpdated 5 years ago
Thank you for your reply. Yes, that's what I ended up doing. My audio clip wasn't silent, but I did a single word. It's a bit rubbish though, because it introduces an unnecessary latency before Nova Sonic starts responding.
Could you provide a bit more exact implementation? I have put a bit of silent audio, but it doesn't work unfortunately
Silent audio does not work, you have to actually put some content in. I just recorded the word "talk" and played that. It only took a few buffers and I trimmed the start/end of the audio to minimise it. It is by far an ideal solution.
If I was in charge of the Bedrock Runtime / Nova roadmap, I would have an option to avoid having to do this....