FastAPI Flux Text-to-Speech

Get started using Deepgram's Flux streaming text-to-speech (v2 speak) with this FastAPI demo app.

Unlike the Flux transcription starter (fastapi-flux), which is a raw websockets proxy, this starter uses the official deepgram-sdk async AsyncDeepgramClient + speak.v2 API — the cleanest fit for an async streaming bridge.

How it works

Browser  ──(JSON: Speak/Flush/Close)──▶  FastAPI backend  ──(SDK speak.v2)──▶  Deepgram Flux TTS
Browser  ◀──(binary audio + JSON control)──  FastAPI backend  ◀──(SDK speak.v2)──  Deepgram Flux TTS

• The API key stays server-side; the browser authenticates with a short-lived JWT (/api/session) on the access_token.<jwt> WebSocket subprotocol.
• A background task iterates async for message in connection and forwards audio (bytes) to the browser as binary and control messages as JSON; the main loop forwards browser Speak/Flush/Close via await connection.send_*.

Client → backend message protocol

Send JSON text frames on the /api/tts WebSocket:

{ "type": "Speak", "text": "Hello from Flux." }
{ "type": "Flush" }
{ "type": "Close" }

Prerequisites

• Python 3.10+
• A Deepgram API key (console.deepgram.com)

Local Development

make init
cp sample.env .env   # add your DEEPGRAM_API_KEY
make start           # backend on http://localhost:8081

Configuration

Variable	Default	Description
DEEPGRAM_API_KEY	—	Required. Your Deepgram API key.
PORT	8081	Backend port.
HOST	0.0.0.0	Backend host.
DEEPGRAM_TTS_MODEL	flux-alexis-en	Flux voice (flux-{voice}-{language}).
SESSION_SECRET	random per boot	Set in production for stable JWT signing.

model, encoding, and sample_rate may also be passed as query params on /api/tts.

License

MIT - See LICENSE