Transcribing an Hour of Calls Now Costs About 30 Cents: Run the One-Week Audio Audit

What shipped. Google moved Gemini 3.5 Transcribe into public preview on August 26. It replaces Chirp 3 as Google's flagship speech-to-text and it is the first Google transcription model sold as a Gemini model through the same API developers already use for text, in a recorded-audio version and a live streaming version. Google reports an average word error rate of 2.6% on recorded audio and 4.0% streaming, measured by Artificial Analysis, with automatic language detection across more than 85 languages including a speaker who switches languages mid-sentence.

The part that matters to an operator is what it does past the raw text. It strips filler words, resolves self-corrections so a customer who says "Tuesday, sorry, Wednesday" comes out as Wednesday, labels who said what, and formats the output. Twelve months ago that was a second cleanup pass you paid for separately.

The price. Third-party pricing trackers put the recorded model near half a cent per minute of audio at list, and the live version near a cent, with a free tier during preview. Round it to 30 cents an hour. Forty hours of sales calls in a month becomes searchable text for about twelve dollars.

The so-what. Most operators will never touch an API. This still reaches your desk inside a quarter: Google is already putting the model behind Gboard dictation and bringing it to Chrome, and every notetaker, phone system, and CRM that resells transcription now prices against half a cent a minute. So the excuse expired. Your estimate calls, your inbound inquiry calls, and your tech-to-customer conversations are the highest-density business data you own, and most owners still run on memory of them.

The move: the one-week audio audit.

1. Pick one conversation type that decides money. Inbound inquiry calls, estimate or quote calls, the doorstep conversation, discovery calls. One type only. Mixing types ruins the pattern.
2. Check your recording rules before you record anything. Consent law varies by state, some states require every party to agree, and some industries carry their own rules on top. Add the disclosure sentence at the start of the call and log that you said it. If you are unsure how your state treats it, ask your attorney before step 3. This step is not optional.
3. Record one week. Use what you already have: your phone system's recording, a meeting notetaker, a call recorder app. Ten to twenty conversations is enough to see a pattern.
4. Get them into text. Whatever transcription your existing tools include is fine. Accuracy matters less than having every word in a file you can paste.
5. Strip identity before you mine. First names or roles, no last names, no phone numbers, no addresses, no card or account details. Run the mining pass in a private or business AI workspace with training turned off.
6. Run the mining prompt below once, on all of them at once. The value is in the pattern across twenty calls, not in any single call.
7. Change one thing, then repeat in 30 days. One script change, one FAQ answer, one price explanation.

The mining prompt:

Role: You are a revenue operations analyst reviewing recorded customer conversations for a [industry] business.

Context: Below are [N] transcripts of [call type] from one week, with names and contact details removed. What I sell and my price range: [one line]. What happens after this call in my process: [one line]. How many of these turn into paying customers: [your rough close rate].

Task: (1) Every question a customer asked, grouped and ranked by how often it came up, with counts. (2) Every objection or hesitation, in the customer's exact words, ranked by frequency, with a note on where in the call it appeared. (3) The turn in each conversation: the sentence right before the customer got interested, and the sentence right before they went cold. (4) Three things my side says that are working, quoted. (5) Three things my side says that cost us, quoted, each with a replacement line I can read out loud. (6) The five answers I should publish on my website based on what people actually asked. (7) One change to test next week and the single number that would prove it worked.

Format: Seven numbered sections in that order. Quote the transcripts rather than paraphrasing. Put counts next to every ranked item.

Constraints: Use only what is in the transcripts and invent no quotes. If a full name, phone number, address, or payment detail slipped through, flag it and keep it out of your output. Do not score or rank individual employees by name. This is a review of the script, so keep the findings about language and process.

What it's worth. Twenty conversations you already had, mined once, tell you the objection you answer badly and the question your website should have answered before the phone rang. The recording week costs you nothing you were not already doing. Budget an hour for the mining pass and one afternoon for the fix.