Tools · Voice · 10

Turn call traffic into monthly cost and required capacity.

Separate connected minutes, audio sent to STT, synthesized audio, generated characters and model tokens. Then size workers and provider limits without confusing simultaneous calls with simultaneous TTS requests.

Costtelephony + media + STT + LLM + TTS

Volumeminutes and billable units by component

Capacitypeak calls, workers and provider concurrency

Traffic and usage

Calls per month

Average duration (min)

Billable STT audio (%)Share of connected time you actually send/bill as STT audio. Use 100% if you continuously stream the caller track.

Synthesized agent audio (%)Share of connected time played as TTS audio. It may overlap user audio during barge-in.

Service hours per monthTurns monthly volume into average concurrency.

Peak simultaneous callsUse traces or a forecast; a monthly average cannot recover the peak.

Capacity

Target worker utilization (%)

Maximum sessions per workerMeasured capacity of your runtime, not a universal limit.

STT sessions per active call1 = one persistent STT session per call. Change it if you multiplex, share or open more than one.

TTS generation duty (%)Share of peak calls generating TTS at the same instant. This is not the share of time audio is playing.

STT concurrency limit0 = do not check. Enter the real account/provider limit if you have one.

TTS concurrency limit0 = do not check. For WebSockets, follow your provider's documented concurrency semantics.

Usage rates

Telephony (USD/connected min)

Media Stream (USD/connected min)

STT (USD/audio min)

TTS (USD/1,000 characters)

Characters per synthesized minuteEditable assumption for converting agent audio minutes into TTS characters.

Other fixed cost (USD/call)

LLM per call

Input tokens

Output tokens

Input (USD/MTok)

Output (USD/MTok)

Monthly cost——

Cost per call—includes every modelled line item

Cost per connected minute—comparable across different call lengths

Connected minutes—

Billable STT audio—

Synthesized audio—

Telephony

——

Media Streams

——

STT

——

TTS

——

LLM

——

Fixed per call

——

Average concurrency—Peak/average: —

Workers required—target / physical sessions: —

STT sessions at peak——

TTS requests generating at peak——

Reference snapshotSpain reference · cascade

Public rates and assumptions verified on 2026-08-21; edit them to match your contract, region and architecture.

Data: 2026-08-21

Method

Cost and capacity do not share one unit.

Billing. Telephony and Media Streams follow connected time; STT depends on the audio you actually process; TTS may be billed by generated characters; the LLM by tokens. The STT-audio percentage should match the real pipeline: use 100% when you stream the full caller track, even through silence.

monthly_cost = connected_min × (telephony + media) + stt_audio_min × STT + agent_chars/1000 × TTS + LLM_tokens + fixed_per_call

Call concurrency. Average concurrency follows from volume and service hours. Peak concurrency must come from traces or a forecast. Sizing from the mean alone is unsafe when call arrivals are bursty.

avg_concurrency = connected_minutes / service_minutes

workers = ceil(peak_calls / (sessions_per_worker × target_utilization))

STT. In a typical cascade, each active call keeps a persistent transcription session, so the reference uses one STT session per call. It is an editable architecture assumption. Twilio documents one WebSocket connection per Media Stream and one bidirectional stream per Call.

stt_concurrency ≈ peak_calls × stt_sessions_per_call

TTS. Do not use audio playback share as provider concurrency. ElevenLabs documents that with WebSockets only time spent generating audio counts toward concurrency and says a limit of 5 can typically support about 100 balanced voice-agent conversations. The initial 5% factor reproduces that heuristic (5/100); it is editable and does not replace p95/p99 measurements from your own requests.

tts_concurrency ≈ peak_calls × generation_duty

Reference rates. On 2026-08-21: Twilio Spain lists $0.0178/min for a local outbound call and $0.0044/min for Media Streams; OpenAI lists GPT Live Transcribe at $0.017/audio-minute and GPT-5.6 Luna at $0.20/$1.20 per MTok input/output; ElevenLabs lists Flash/Turbo TTS at $0.05 per 1,000 characters. Every rate remains editable.

Not an invoice or SLA. The model excludes monthly phone numbers, taxes, recording, observability, tool calls, storage, transfers, volume discounts and contractual minimums. Provider concurrency limits depend on plan and account and must come from your real configuration.

To optimize response time as well as cost, use the voice-agent latency explorer. For architecture trade-offs, read three architectures for voice agents.

Sources: Twilio Voice Spain · Twilio Media Streams · OpenAI GPT Live Transcribe · OpenAI GPT-5.6 Luna · ElevenLabs API pricing · ElevenLabs concurrency. Verified 2026-08-21.