Voice Assistant
Architectureaiserverless
Connect + Lex voice loop with Bedrock fallback and session memory
Open & fork this on Zstem →Architecture
- Voice Channel: Audio in and out of the conversation
- Understanding: Turning speech into intent
- Fulfillment: Acting on the intent
- Caller: Phone or in-app voice user
- Amazon Connect: Contact-center entry point; streams audio both directions
- Transcribe: Streaming speech-to-text for the caller's audio
- Polly: Neural text-to-speech for the assistant's replies
- Lex Bot: Intent + slot recognition over the transcript
- Bedrock Fallback: Handles low-confidence turns conversationally instead of failing
- Fulfillment Fn: Executes the resolved intent against backend systems
- Session Store: Dialogue state and slot values across turns
- Call Analytics: Intent match rates, containment and latency metrics
Sequence
Participants: Caller, Amazon Connect, Transcribe, Lex Bot, Fulfillment λ, Session Store, Polly
- Caller→Amazon Connectaudio stream
- Amazon Connect→Transcribestream audio
- Transcribe→Amazon Connectlive transcript
- Amazon Connect→Lex Bottext turn
- Lex Bot→Fulfillment λintent + slots
- Fulfillment λ→Session Storeload/update session
- Session Store→Fulfillment λsession state
- Fulfillment λ→Lex Botfulfilment result
- Lex Bot→Amazon Connectresponse text
- Amazon Connect→Pollysynthesize
- Polly→Amazon Connectaudio
- Amazon Connect→Callerspeech reply