Skip to the content

Case study

An AI service billed per second also charges for silence

Every second sent to a voice service is billed, including the quiet after the last word. On a client's tool, roughly a third of each clip was that silence. Here is how it was traced, trimmed and tested.

Key takeaways

  • A bill by the second counts every second you send, not every second your customer hears.
  • The quiet after the last word is the easiest waste to find, because no finished edit ever plays it.
  • Shorten a clip only when the speech end has been measured and the audio itself confirms it, and send everything else whole.
  • Compare the old result with the new one on real examples before you treat a lower bill as a safe one.

If you pay an AI service billed per second, the meter runs on everything you send it, including the quiet after the last word that no listener ever reaches. On a client's video tool, roughly a third of every clip sent away for a voice change was exactly that: silence. We traced where it came from, changed what gets sent, tested the result against whole clips, and had it live within one working day; a voice change on that platform now costs roughly a third less per clip.

What a per-second bill hides

A per-second rate tells you the price of one second, not how many of the seconds you pay for ever reach your customer.

The bill arrives as a count of seconds used. It does not separate the seconds that carried speech from the hush before someone starts talking or after they stop, and a single total hides that difference well. On a tool that handles clips all day, it is easy to skim past.

Some of that quiet is unavoidable: a pause in the middle of a sentence is part of the performance and should stay. The part worth hunting is the tail after the last word, which the finished piece cuts away anyway.

Why the waste happens in tools like this

Sending a clip whole is the simplest thing to build, so in our experience it is what gets built.

The service receives the file from beginning to end and bills every second of it, room tone included. The finished edit, though, cuts a fraction of a second after the speech ends, so none of that tail is ever played.

The two lengths live in different places. The bill counts seconds received; the edit counts seconds kept. Until somebody puts one beside the other, the gap stays out of sight.

How our engineer traced it

The tool already knew where the speech ended in every clip, because that is the point where the edit is cut. That measurement was the thread our engineer pulled.

The next question was how much of what follows that point is actually played. Set against the two kinds of edit in the tool, it came to a little over a tenth of a second in one and under a third of a second in the other. Everything sent after it was paid for and thrown away, and on these clips that came to about a third of every one.

The obvious worry was quality. A voice service works from the clip it is given, and a shorter clip might come back sounding different at the edges. So the change was not treated as finished until it had been run on real clips with real conversions and compared against the old way of sending the whole thing.

What we changed

Each clip is now sent only up to half a second after its last word, and never cut closer than four tenths of a second past that point.

Two checks have to agree before a clip is shortened. The speech end must have been measured on that very clip, and the clip's own sound must confirm that the speech is over — that the tail is quiet, not a held note or the start of another word. When both agree, the changed voice comes back and is padded with silence to the clip's exact original length, on our side, after the service has returned it. That padding is never sent, so it is never billed, and the edit receives a clip just as long as before.

Anything that fails either check is sent whole, as it always was, and the log records why.

What you would see if this were your tool: a shortened clip is billed for fewer seconds than its own length, and the log carries a reason for it. A clip sent whole is billed for exactly its length, with the reason beside it in the log — the speech end could not be measured on that clip, or the sound at the tail did not read as quiet.

What the tests showed, and what is still whole

The change went live on the client's platform after tests on real clips with real conversions, and it behaves as the old way did within the limits we could measure.

In those tests the end of the speech moved by a few thousandths of a second at most. The shape of the voice matched as closely as two whole conversions of the same clip match each other, and loudness stayed inside the normal difference between two runs of the same clip. That is a comparison, not a promise: it says we could not measure a change, not that none exists.

One part of the tool is untouched. Changing the voice of a whole ad at once does not yet use the measurement and still sends whole clips.

What you can check yourself

  • Take a clip you have already had processed and set its own length beside the seconds billed for it. If the two are the same, that clip went whole.
  • Find the tool's log or history. A shortened clip should have a reason recorded next to it, and so should a clip that was left alone.
  • Ask whether every job in the tool uses the shorter send or only some. In this tool, changing the voice of a whole ad at once still sends whole clips.
  • Ask what happens to audio the tool cannot judge safely. The answer to look for is that it goes through as before, whole.
  • For a service billed by the minute, ask whether long quiet stretches are sent at all, and whether anything uses them.

What to do next

Start with one number, on one kind of job: how many seconds you send for processing, against how many end up in the finished work. If nobody can show you that on real examples, leave the tool as it is and ask again before anything is trimmed. Trimming what a paid service receives, and proving the output still sounds the same, is the kind of work our engineers do for clients.

Questions people ask

How do I know if I am paying for silence in an AI voice or video tool?

Compare two lengths: what you send to the service, and what ends up in the finished piece. If a clip is sent whole and the edit cuts a fraction of a second after the last word, the tail is being billed and never played. The bill alone will not show this, because it only reports seconds used.

Will trimming the clip change how the voice sounds?

In our tests on real clips it did not, within what we could measure. The end of the speech moved by a few thousandths of a second at most, the voice matched as closely as two whole runs of the same clip match each other, and loudness stayed inside the normal difference between runs. The trim is also only made when two separate checks agree, and the changed voice is padded back to the original length.

What happens to clips the tool cannot judge safely?

They are sent whole, exactly as before. A clip is shortened only when the end of the speech was measured on that clip and the clip's own sound confirms the speech is over. The tool also records why each clip was shortened or left alone.

Does this apply to transcription and video services too, or only voice?

We only did this for voice changes on one client's video tool, so we cannot tell you what it would save elsewhere. The question is worth asking of anything billed by the second or minute: transcription of long recordings with quiet stretches, or processing that pads clips. Ask where the unused seconds go.

How much of a clip is usually silence?

In the clips we looked at, about a third of each one was quiet after the last word. That comes from one tool and one set of clips, so treat it as a reason to measure your own rather than as an expectation. The share will differ by presenter, script and how the recording was made.

From Tech Concierge

Want this handled for your business?

Your own engineer builds and runs the website, the app, the CRM and the AI assistant that answers your customers.

The first 7 days are free.

Read next

Case studyVideo and content

Why your video tool promises a day for a job ready by evening

A finish time built on one fixed rate is furthest from the truth on the biggest jobs, while the counter beside it can sit at zero. What we found on a client's video platform, and what to check on yours.

5 min read

Case studyVideo and content

Why your tool calls a long job stuck when it is still running

Twice in one week a video platform told a team an ad had stopped while it was still being made. The fix was to ask the step holding the job before calling it lost, instead of deciding from a clock.

5 min read

Case studyVideo and content

Why your AI presenter changes background mid video

A client's video team saw their AI presenter change place halfway through two new ads, and remaking the lines did not help. The cause was a setting that had been switched off in one kind of video only.

4 min read

We count visits to this blog with our own analytics, never Google's, and share what you read with nobody. May we remember this browser with a cookie? See the privacy policy.