A Checklist Before You Spend Speech-to-Text Credits
Speech-to-text bills by audio minute, not by usable transcript. A checklist to run before the first paid minute, and what to do when the wallet is already low.
Why speech-to-text credits vanish on the wrong file
Speech-to-text bills by audio minute. A 60-minute recording looks cheap until you run it three times. Credit wallets of 10, 30 or 120 minutes empty overnight when the first pass used the wrong model, the wrong language hint, or a file that was already too damaged to transcribe. Public price sheets such as Google Cloud Speech-to-Text pricing make it plain: you pay for audio processed, not for a transcript you can use.
The usual failure is not a bad model. It is audio that cannot produce a usable transcript however many credits you spend. Silence-padded exports, stereo calls with two people on one channel, 8 kHz phone audio upsampled to 16 kHz, and video files with a tiny audio track all look valid in a file picker. The meter still moves.
Checklist before the first paid minute
Measure duration, not file size. A 200 MB video can be 4 minutes; a 15 MB WAV can be 90. Credits follow duration.
Trim leading and trailing silence. Two minutes of hold music on a 20-minute call is about 10 percent of the bill and zero words.
Check sample rate and channels. Most models expect 16 kHz mono. Sending 48 kHz stereo does not buy accuracy, and many vendors resample and still bill the original length. Downmix to mono yourself so you know which channel you kept.
Listen to 20 seconds at the start, the middle and the end. If you cannot hear the speaker, the model will not either. Music beds, heavy noise gates and clipped peaks produce confident wrong words, and re-encoding will not fix clipping already in the waveform.
Trial on 60 seconds, not the whole file. If names and numbers are wrong on that slice, trying a second model on the full recording multiplies the cost: a 45-minute file tried on three models is 135 billed minutes. Write down the balance before and after the one-minute test, so the real burn rate is visible.
Read what "credits" includes. Some products count uploaded minutes, some successful minutes, some every retry. If a stalled job still takes from the wallet, cancel it before starting another.
Split "good enough to search" from "good enough to quote". A skim can use a cheaper model. A transcript you will publish needs a human pass on names, numbers and timestamps.
Store the settings with the output: model, language, minutes billed and the balance after. Without that note the next person runs the same file again and pays twice.
When the wallet is already low
If fewer than 15 minutes remain, do not start a longer file. Split at chapter markers and transcribe only the part you will use. Export a clean mono M4A or WAV first, so the job does not fail on a container it cannot open. A failed job that still bills is the most expensive kind of silence.
From the author, JK33 Jin: the ClipShrink video-to-text converter shows the duration and credit cost of a file before you transcribe it.
