Nova Sonic
Lira uses Amazon Bedrock Nova Sonic (amazon.nova-sonic-v1:0) as the core speech AI engine providing bidirectional speech-to-speech processing.
Capabilities
Nova Sonic combines three operations in a single bidirectional stream:
- Speech-to-Text — Transcribes incoming audio in real-time
- LLM Reasoning — Processes the transcription with context (system prompt, org knowledge base, meeting history)
- Text-to-Speech — Generates natural-sounding voice responses
Configuration
| Variable | Value | Description |
|---|---|---|
LIRA_BEDROCK_REGION | us-east-1 | AWS region for Bedrock API calls |
LIRA_NOVA_SONIC_MODEL_ID | amazon.nova-sonic-v1:0 | Model identifier |
Context Injection
Before Nova Sonic processes a response, lira-context-builder.service.ts assembles relevant context:
- Organization knowledge base — Vector search results from Qdrant
- Meeting transcript — Recent conversation history with speaker attribution
- Personality mode — System prompt variation (Supportive, Challenger, Facilitator, Analyst)
- Integration data — Relevant data from connected tools (Linear issues, calendar events, etc.)
Session Management
- Keepalive — Silent audio frames are sent periodically to prevent session timeout
- Barge-in — Nova Sonic stops generating audio when new input is detected mid-response
- Error recovery — On connection failure, the bot logs the error and attempts to re-establish the stream