The Voice Stack — overview

One Floe key pays every leg of a voice conversation — telephony, speech-to-text, LLM, and text-to-speech — and every leg lands in one ledger with one set of budget caps.

What one conversation costs

A single spoken turn touches four vendors. With Floe, all four bill to the same balance, and one task budget caps the whole turn. Costs below are illustrative.

Leg	Vendor	Typical cost	Status
Telephony	Twilio	~$0.014 / min	Live
Speech-to-Text	Deepgram	~$0.007 / min	Live
Reasoning	LLM (any)	~$0.020 / turn	Live
Text-to-Speech	ElevenLabs	~$0.007 / turn	Live
Total		~$0.048

Run a full voice turn

Every leg carries the same X-Floe-Task-Id: call-8842, so one task budget covers the whole conversation. Each paid response returns the cost of that leg in X-Floe-Payment-Amount.

1. Transcribe (Deepgram, via the proxy)

curl -X POST https://credit-api.floelabs.xyz/v1/proxy/fetch \
  -H "Authorization: Bearer $FLOE_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Floe-Task-Id: call-8842" \
  -d '{"url": "<DEEPGRAM_STT_ENDPOINT>", "method": "POST", "body": "{\"audioUrl\":\"https://example.com/turn.wav\"}"}'

2. Reason (LLM, OpenAI-compatible)

curl -X POST https://credit-api.floelabs.xyz/v1/chat/completions \
  -H "Authorization: Bearer $FLOE_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Floe-Task-Id: call-8842" \
  -d '{"model": "openai/gpt-4o", "messages": [{"role": "user", "content": "Caller asked to reschedule to Friday. Confirm."}]}'

Just the Floe key — no OpenAI key. Floe holds the provider credential and bills the tokens to the same balance as the other legs. To keep your own provider key instead, use the BYOK endpoint — see Add Floe to your existing pipeline.

3. Speak (ElevenLabs, via the proxy)

curl -X POST https://credit-api.floelabs.xyz/v1/proxy/fetch \
  -H "Authorization: Bearer $FLOE_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Floe-Task-Id: call-8842" \
  -d '{"url": "<ELEVENLABS_TTS_ENDPOINT>", "method": "POST", "body": "{\"text\":\"You are booked for Friday at 2pm.\"}"}'

All three legs share X-Floe-Task-Id: call-8842, so a single task budget caps the whole conversation — telephony, STT, LLM, and TTS together. A token router only sees step 2. Floe sees steps 1, 2, and 3 on one bill.

Live voice vendors

The Voice vendor directory is the canonical roster of live TTS, STT, and telephony vendors — with endpoints, request shapes, and per-vendor pricing.

Live voice with your own stack (LiveKit / Pipecat)

The turn above is batch — each leg is a discrete request/response call. If you're running a live conversation with LiveKit or Pipecat, the framework streams audio in and expects an STT plugin that emits interim/final transcripts as the caller speaks. All three legs run on Floe keyless, on one ledger:

Leg	On Floe?	How
LLM	Yes — keyless	Point the framework's LLM service base_url at https://credit-api.floelabs.xyz/v1 with your Floe agent key. OpenAI-compatible; metered on Floe.
TTS	Yes — keyless	Floe's POST /v1/audio/speech (OpenAI-compatible), or any vendor via POST /v1/proxy/fetch. Metered on Floe.
Live STT	Yes — keyless	Open a WebSocket to wss://credit-api.floelabs.xyz/v1/audio/transcriptions/stream?model=deepgram/nova-3&encoding=linear16&sample_rate=16000&language=en with your Floe agent key. Stream raw PCM frames up; Floe streams {type:"transcript", text, is_final, speech_final} events down (is_final:false = interim, true = final; speech_final:true = end of utterance) — the interim/final feed a LiveKit/Pipecat STT plugin consumes. Metered per audio-second on Floe — Floe fronts the Deepgram key.

So a live stack is Floe for the LLM, TTS, and STT legs — keyless, one ledger, one set of budget caps. See Floe Inference — streaming transcription for the wire protocol (frame encoding, sample-rate bounds, message shapes).

If you don't need a self-hosted LiveKit/Pipecat pipeline, the keyless realtime WebSocket (wss://credit-api.floelabs.xyz/v1/realtime?model=<provider/model>) is a different tool: a speech-to-speech model (openai/gpt-realtime, google/gemini-live-3.1, xai/grok-voice) that takes audio in and returns audio out on one connection, metered on Floe. It is not a drop-in streaming-STT feed — it won't emit interim/final transcripts into a LiveKit/Pipecat STT plugin. Reach for it when you want an end-to-end voice model; reach for the streaming-transcription endpoint above when you need the STT leg of a BYO stack.

One key, one ledger

Because the LLM leg routes through Floe's OpenAI-compatible endpoint, reasoning tokens land on the same ledger and the same budgets as the vendor legs — not a separate provider bill. See Unified Billing & Ledger to read the combined statement, and Spend Controls to set the task, vendor, and session caps that govern every leg.