Research-backed calculator guide

Does Silence Cost Money in Voice AI? Model Dead Air, Hold Time and Listening Meters

Learn when voice AI silence, hold time, and dead air are billed, how provider meters differ, and how to calculate the effective cost of a productive minute.

Why the meter matters more than the word “silence”

A live voice call usually contains several clocks. The carrier clock starts when the call is connected. A voice-agent platform may use that connected duration for its bundled rate. An STT stream may begin when audio is sent and end when the WebSocket closes. The LLM clock is token-based and changes with prompts, history, and tool calls. TTS may be priced by characters, audio seconds, or a bundled voice minute. These clocks can overlap without producing one shared definition of a minute.

That is why “does speech to text charge for silence?” has two answers. If the STT service receives a continuous stream and charges the duration of that stream, quiet audio or a pause can remain part of the billable interval. If it charges only a provider-defined processed-audio quantity or a trimmed segment, the answer can be different. Do not substitute a voice activity detector's label for the provider's invoice field.

There is also an operational cost. A long end-of-turn timeout can add dead air before the agent speaks. Even when that delay does not create a separate STT line item, it can lengthen the connected call, increase the platform or carrier minutes, and cause a caller to repeat the request. The right analysis therefore separates “silence billed by the meter” from “silence that makes the call longer.” Both can affect unit economics.

Worked cost example

Use this example as an estimate, not a vendor quote. Assume 1,000 U.S. calls per month, each connected for 6 minutes. Assume 45 seconds of each call is silence, hold, or an avoidable response gap. Use a hypothetical bundled rate of $0.10 per connected minute. The rate is deliberately labeled as an input so you can replace it with your contract or rate card.

Under this connected-time assumption, silence is 12.5% of the connected minutes and $75 of the modeled bill. The effective productive-minute rate is 14.3% higher than the nominal $0.10 rate. The arithmetic is not claiming that every provider identifies a $75 “silence” line item. It allocates the connected-time bill to show the economic effect of the assumption.

If the same stack instead billed only the 5,250 minutes that passed its documented audio or speech meter, the modeled bill would be $525 and the effective productive-minute rate would remain $0.10. That is a scenario, not permission to assume silence trimming. Confirm the rule, the rounding increment, channel treatment, and session close behavior in the vendor documentation and your usage export.

Retell's pricing page displayed a $0.11/minute example configuration when checked on 2026-11-02. Substituting that observed vendor calculator output into the connected-time scenario gives 6,000 × $0.11 = $660, of which 750 × $0.11 = $82.50 is the allocated silence/hold portion. Treat this as a dated vendor-page observation, not a promise that every Retell configuration costs $0.11/minute. ^[1]

A practical model for dead air, hold time, and listening meters

Start with an event ledger rather than a single stopwatch. For each call, capture connected_at, connected_at_end, audio_stream_started, audio_stream_ended, speech_seconds, silence_seconds, hold_seconds, transfer_at, and the provider-reported usage quantity. A transfer deserves special care: Retell says its AI-agent fee stops after transfer while telephony continues, so the same call can have different end times for different meters. ^{[1] [2]}

Next, classify the meter into one of four useful types:

1. Connected-call meter. The bill follows the phone connection or bundled agent duration. Silence and hold usually remain inside the interval unless the provider documents an exclusion.
2. Session or WebSocket meter. The bill follows the open realtime session. A late close, reconnect, or idle socket can add cost even when no one is speaking. Deepgram explicitly describes its managed Voice Agent pricing as calculated on WebSocket connection time. ^[3]
3. Processed-audio meter. The bill follows audio duration received or processed, often at per-second precision. Deepgram's STT page publishes per-second billing and says normal audio requests do not use block-based rounding. AssemblyAI describes Universal-3.6 Pro Realtime at $0.45/hour, billed per second, as checked in its October 2026 article. ^{[3] [4]}
4. Speech or trimmed meter. The provider excludes some non-speech or applies endpointing before rating. Never infer this from a dashboard label such as “speech minutes”; find the definition and test a controlled file or call.

These categories are an analysis framework, not a claim that every vendor uses one pure meter. A platform can bundle a connected-time agent charge, a carrier charge, an STT audio charge, and a token-based LLM charge on the same call. Show them as separate rows so an apparently cheap provider rate does not hide an unmodeled line.

For U.S. telephony, carrier minutes can remain payable even when the AI component changes. Twilio's U.S. pricing page, current there as of August 2026, lists pay-as-you-go local outbound calls at $0.0140/minute and local inbound calls at $0.0085/minute, with separate feature charges such as recording and transcription. Those are carrier and feature rates, not a complete voice-agent cost. ^[5]

How to audit your own invoice

First, choose one call that contains a clear hold or quiet interval. Export the call's raw connected duration, the platform's billed duration, and each underlying provider's usage record. Second, compare timestamps rather than transcript word counts. A transcript can omit silence while the session remains open. Third, repeat the test with a deliberately short pause and a clean hang-up. This reveals minimums, rounding, and delayed session closure.

Fourth, calculate variance in both minutes and dollars. A simple reconciliation is provider billed minutes − internal expected minutes, followed by variance × rate. If your internal ledger says 5,250 minutes but a connected-time provider bills 6,000, the 750-minute difference is not automatically an error. It may be the documented treatment of hold, silence, transfers, or rounding. If the rule does not explain it, flag the row for investigation.

The AI API invoice reconciliation calculator is useful when you need to turn those usage rows into a repeatable variance check. For the other side of the comparison, use the model pricing comparison route to keep token, audio, character, and minute units separate. If dead air is a quality problem as well as a billing problem, the AI observability ROI calculator can help model the value of measuring response gaps, retries, and abandonment assumptions.

FAQ

Is silence billed in voice AI?

It can be. A provider that meters the connected call, active voice-agent session, or continuously open STT stream may include silence and hold time. A provider that bills a documented trimmed or speech-only quantity may not. Read the meter definition and test it with usage data; “per minute” alone is not evidence.

Does hold time cost the same as spoken time?

It costs the same when both fall inside one connected-time or session-time interval, because the meter is measuring elapsed time rather than speech content. A transfer can change the answer if the AI fee stops at handoff while the carrier continues. Model AI, STT, TTS, LLM, and telephony end times independently.

Does speech to text charge for silence?

There is no universal answer. Deepgram publishes per-second audio pricing for STT and separately documents WebSocket connection-time pricing for its Voice Agent product. Other stacks can trim, endpoint, or bundle audio differently. Ask whether silence is included in received audio, processed audio, or connection duration, and whether each channel is rated separately. ^[3]

How do I calculate voice bot dead-air cost?

Multiply the silence or hold minutes you are allocating to the relevant meter by that meter's rate. In the worked example, 1,000 calls × 0.75 minutes of silence × $0.10/minute equals $75. Then divide the total bill by productive minutes to get the effective productive-minute rate. Keep this as an allocation unless the invoice exposes silence as a separate measured quantity.

What should I ask a provider before comparing per-minute prices?

Ask what starts and stops the meter, whether connected silence and hold are included, how transfers and failed calls are handled, whether a minimum or rounding increment applies, whether multichannel audio is multiplied, and when an idle WebSocket closes. Also ask which costs are bundled and which remain separate, including telephony, LLM tokens, TTS, recording, concurrency, and add-ons.

Assumptions and limitations

This guide uses a dated research snapshot from 2026-11-02 and English/U.S. framing. Vendor pages, rates, model names, regional telephony charges, and billing exceptions can change. The $0.10 worked example is an explicit estimate; the $0.11 Retell figure is an observed vendor calculator configuration, not a universal quote. The example allocates silence cost from elapsed time and does not prove that a provider's invoice has a silence line item.

The calculation excludes taxes, discounts, committed-use pricing, concurrency fees, phone-number rentals, human transfer costs, retries, failed-session charges, transcription storage, and LLM/TTS usage unless stated. It also assumes one billed connected interval per call. Replace those assumptions with exported usage rows before using the result for pricing or margin decisions.