Turn recorded speech into text you can trust

Audio transcription and audio conversion solve different problems. Conversion changes the file format. Transcription turns the speech inside the file into editable text.

To transcribe an audio recording well, create an editable first pass, keep timestamps and speaker context, then check the words most likely to change the meaning: names, numbers, dates, negations, decisions and action owners. Before choosing a service, also decide where the recording may be processed and how long it may be retained.

Choose the required output before processing the recording. A verbatim record may preserve false starts and filler words. Meeting notes may need decisions and assigned work. Interview material may need precise quotations, and the source recording should remain unchanged so disputed wording can be checked against it.

Processing location belongs in that decision. Check where the audio goes, whether another copy is stored and when that copy is deleted. A product’s description of privacy is less useful than its capture path, network requests and retention terms. Our account of local-first architecture explains the questions to ask.

Make the first pass reversible

Use clean source audio where possible. Google’s speech recognition guidance identifies close microphones, low background noise and balanced speaker volume as helpful conditions. Overlapping voices, unfamiliar proper names and specialist terms make recognition harder. Converting a compressed file into another format does not restore detail that the original recording lost.

Upload services suit an existing file when sending it to a provider is permitted. Free access has three separate meters: per-file duration, monthly minutes and lifetime imports. Otter Basic provides 300 minutes per month, caps each transcription at 30 minutes and permits three file imports across the account’s lifetime, checked September 2026. TurboScribe provides three free files per day with a 30-minute maximum for each file, checked September 2026.

ChatGPT Record is a better fit for eligible users who want to make the recording inside its macOS desktop app and continue working with the result in ChatGPT. It is available to Plus, Enterprise, Edu, Business and Pro workspaces, checked September 2026. Selecting Send after recording uploads the transcript and summary. It is an in-app recording workflow rather than a general local file converter, and OpenAI warns that its transcriptions may contain mistakes.

For a format change without speech recognition, FFmpeg is the better tool. It records, converts and streams media and is distributed under the LGPL with optional GPL-covered components, checked September 2026. Manual work avoids automated recognition but takes time: Oxford estimates four to six hours for one clear hour of interviews, lectures or podcasts and six to ten hours for meetings or group discussions.

Audit the words that can change the record

Treat machine output as a draft. Google describes insertion, substitution and deletion as the main word-error types. A polished paragraph can still contain a substituted surname, a missing negative or an inserted word that reverses the intended meaning. Compare uncertain passages with the recording instead of correcting them from memory.

Check names first by searching for each participant, company, product and place mentioned in the recording. Then inspect numbers and dates beside their original timestamps. Pay particular attention to prices, percentages, quantities, version numbers, deadlines and calendar dates. Similar-sounding digits can produce grammatically plausible sentences that remain factually wrong.

Search for negative constructions, including not, no, cannot and contractions, then listen to the surrounding passage when one appears or seems to be missing. Mark each sentence that records a decision. Confirm what was approved, rejected or deferred. A summary that preserves the topic but reverses the decision is not a usable record.

Review speaker attribution beside decisions and commitments. Confirm who proposed an option, who accepted it and who owns the follow-up. Noats separates voices within each meeting and lets the user assign names during or after the call. The speaker-separation workflow explains how those meeting-specific labels are produced and corrected.

Keep timestamps and speaker labels through the audit. They provide a route from a questionable sentence to the relevant moment and voice. Remove either after review only when the finished document does not need it. Store the corrected result in an editable, searchable format rather than leaving the useful copy inside a playback interface.

Capture the meeting before an upload is needed

A future meeting can skip the later file-upload stage. Noats records both sides on a supported Mac, produces the draft and writes the meeting note in the same local workflow. The two-sided capture guide explains how microphone input and call audio enter that recording.

In Noats, local means the audio is captured, transcribed, separated by speaker, turned into a note and stored on the Mac by default. There is no Noats server that receives meetings. Noats is free during beta and requires Apple silicon with macOS 14.2 or later.

Run this verification on a supported Mac:

1. Start a meeting recording in Noats.
2. Speak through the microphone while remote audio plays through the call.
3. Stop the recording.
4. Confirm that the transcript assigns the microphone channel to your side and separates remote audio into the other speakers.
5. Open the generated note and locate its files in the user’s Application Support folder.

The call itself still travels through the call provider under that provider’s terms. Noats writes the audio, transcript, speaker labels, note, folders and meeting chat to a database and plain files in the user’s Application Support folder. A keystroke exports the note as Markdown for storage and search outside the app.

Run one minute through the method. Include a name, a number, a date, a negation, a decision and an action owner. Keep the draft when all six survive the audit.