It was almost always empty, or it had three bullet points written twenty minutes after the call by whoever was least tired. A week later, engineering would ask, "Did they actually say the import failed, or is that our guess?" and the best we could do was shrug and offer to listen to a 47-minute file.

We were not trying to become a legal transcription department. We just wanted a way to find the sentence again, and audio to text was the least heavy way to get there.

The first change: stop relying on memory

We started uploading the recording the same day — a Zoom m4a, sometimes a phone file from a hallway interview — to an audio to text converter the PM already used for her own interviews. Support for 30+ formats came in handy because our recordings were not consistent. Once, a founder sent a public YouTube walkthrough of their workaround; we pasted the link instead of downloading it.

A one-hour file has usually come back in around two minutes. Noisy office audio still takes longer to clean up on the human side, even if the job itself is fast.

Speaker labels and timestamps are the features we actually kept on. "User" and "us" is enough for a two-person call. When a session has two customers on the line, the split is not perfect, but it is better than a single wall of text.

Then we search: "failed," "workaround," "price," "wouldn't," "need." The interesting moments stop being buried at 0:33:12.

That is the modest promise of audio to text in a product team: not auto-written insights. A record you can search when someone challenges your memory.

What we do with the transcript

Our rule is that the raw transcript never goes into the research repository as the source of truth. It is a draft.

1. The person who joined the call highlights five to ten moments and writes a half-page brief in their own words: what we asked, what we heard, what we will do.
2. Quotes we might show the rest of the company get a timestamp, then a relisten. Tone matters. "Fine, I guess" is not the same as "That's fine."
3. We use an AI summary as a cramped outline only. There are 16 summary styles, and a custom prompt can ask for "decisions and open questions." We still rewrite. A summary once listed a feature request we had suggested to the customer as if they had demanded it.
4. We export Word or TXT into the call folder, next to the brief. PDF and SRT/VTT exist if someone needs them; we do not assume every feature exports every format.

On clear calls, accuracy has been very strong — the product says up to 99% on clear audio. A shared workspace with two laptops open and a dog barking is a different story, and we budget time to correct names and product terms. Vocabulary Groups help when you have a distinctive brand or API name.

Practical limits we ran into

• We signed in for the free plan: 30 credits a day, which can mean up to 30 minutes in Standard mode if you leave speaker identification off, with a cap of three files a day and 30 minutes per file. We usually keep speakers on, and that uses two credits a minute, so we do not treat 30 credits as "always 30 minutes."
• 98 languages and built-in translation helped on one call that drifted between English and Portuguese. We treated the translation as a sketch and checked the original before writing it down.
• Paid plans allow a file up to 10 hours or 5 GB. The free plan does not. We have not needed the high end. Our problem was never file size. It was files nobody opened.

If someone on the team records a quick recap after the call, the in-browser recorder can go through the same audio to text pass. We also do not put customer audio into a tool without a quick consent check. That is a team policy, not a feature of the product.

What got better

The empty "Call notes" doc is gone. In its place is a brief with links to the lines that support it. When engineering asks "did they actually say that?", someone pastes a quote and a timestamp instead of scheduling another call.

If your team is also losing the middle of conversations, you do not need a new process with twelve fields. Record the next interview, turn the customer call into text, and write the brief from a searchable, timestamped transcript.

The recording was never the hard part. Finding the sentence was.