Voice agent¶

STT → LLM → TTS. The middle step is the E2 call-an-llm service imported verbatim — the voice recipe only adds transcription and speech synthesis on the ends.

The pipeline¶

     ┌─────────┐    ┌────────────┐    ┌─────────┐
     │  audio  │──► │  STT       │──► │ prompt  │
     │  bytes  │    │ (cloud)    │    │  text   │
     └─────────┘    └────────────┘    └────┬────┘
                                           │
     ┌─────────┐    ┌────────────┐    ┌────▼────────┐
     │  audio  │◄── │  TTS       │◄── │ E2 chat_*   │
     │  bytes  │    │ (cloud)    │    │ (REUSED)    │
     └─────────┘    └────────────┘    └─────────────┘

Per-cloud services¶

Cloud	STT	TTS
Azure	Azure AI Speech — `SpeechRecognizer` in `azure-cognitiveservices-speech`	Azure AI Speech — `SpeechSynthesizer` (same SDK)
GCP	Speech-to-Text v2 — `google-cloud-speech`	Text-to-Speech — `google-cloud-texttospeech`
AWS	Amazon Transcribe — `boto3` `transcribe` client	Amazon Polly — `boto3` `polly` client

Docs verified 2026-08-08¶

Every voice-id / model / API call below was verified against these live pages:

Azure Speech STT quickstart — learn.microsoft.com/…/speech-service/get-started-speech-to-text
Azure Speech TTS quickstart — learn.microsoft.com/…/speech-service/get-started-text-to-speech
GCP Speech-to-Text v2 — cloud.google.com/speech-to-text/v2/docs/quickstart-client-libraries
GCP Text-to-Speech — cloud.google.com/text-to-speech/docs/quickstart-client-libraries
GCP TTS voice list — cloud.google.com/text-to-speech/docs/list-voices-and-types
AWS Transcribe boto3 — docs.aws.amazon.com/transcribe/…/getting-started-python-sdk
AWS Polly available voices — docs.aws.amazon.com/polly/…/available-voices

Pages¶

STT — audio → text per provider.
LLM in the middle — the two-line reuse of E2 (chat_native / chat_claude).
TTS — text → audio per provider + a verified voice-id table.
Deploy — Terraform composing E1 baseline + E2 call-an-llm; Helm reusing the E2 chart shape.

Validate-only¶

Same discipline as E2/E3 — no live audio, no live LLM, no live TTS in CI. The "verify" block on each page is the local one-liner to test against your own credentials.