The most private place for your session audio is the device already in your hand. You do not need a cloud service, or CoachFlow, to turn a recording into text. With a local-first tool, the audio and the transcript can stay on your own machine.
This is the transcription step, the one that comes before any analysis. Here are the local-first tools I would start with, the cloud options for when local is not practical, and the one distinction that decides how private the whole thing is.
The distinction that matters
Two things get mixed up: where a company is based, and where your data is handled. A tool can be owned by a US company and still process your audio in the EU, or be built in Europe and run on US cloud infrastructure. So the useful split is not the flag on the website, it is where the work happens.
- Local-first means the original audio and transcript can stay on your own device.
- Privacy-mode cloud means a third party still processes the content, though it may offer controls around training, retention, region or deletion.
Local-first is the cleaner position for confidential coaching, because it reduces the number of hands the original audio passes through.
For the record, CoachFlow is not a transcription tool, a meeting bot or a recording store. It starts after the session, on a transcript you have already chosen to use. Most meeting AI starts by capturing more. CoachFlow starts after you have decided what can be used. So the transcription choice is yours, and this post is about making it well.
Local-first tools I would start with
| Tool | Works on | What to know |
|---|---|---|
| Superwhisper | Mac, Windows, iOS | My default. Local voice models, offline use, many languages. Its privacy page says local transcription does not require it to collect your audio or transcripts. |
| MacWhisper | Mac, iOS | Strong for transcribing audio files. Local by default; the optional cloud transcription, translation and AI features send data out. Built by a Netherlands-based developer. |
| Aiko | Mac, iOS | Simple on-device transcription using Whisper. Good when you do not need a full workflow. |
| noScribe | Mac, Windows, Linux | Free and open source, made for sensitive interviews. Runs locally, no cloud. Less polished for everyday use. |
| Buzz | Mac, Windows, Linux | Open source, offline transcription powered by Whisper. Best if you are comfortable with less commercial support. |
Spokenly and VoiceInk are worth a look too, but treat them as local-first only when you have switched on local-only mode, since both also offer cloud features.
One caveat worth stating up front: these tools and their settings change. Before you rely on one, read its current privacy page and check the setting that keeps processing on your device.
A setup for sensitive coaching work
When the recording is a real client session, I keep it simple:
- a local voice model, with no cloud post-processing
- export the transcript to your own device
- remove the identifying details before anything analyses it, following how to remove names from a transcript
The aim is for the audio and the first transcript to stay on your machine, and for only a cleaned transcript to travel any further.
When local-first is not practical
Sometimes you need a cloud service: a long recording, a team workflow, or a language a local model handles poorly. That can be fine, with the right settings. The things to check are the same each time:
- Data residency: can you choose an EU region?
- Retention: is the audio or transcript kept, and for how long?
- Training: can you opt out of your data improving the provider’s models?
- Deletion: can you remove the data when you are done?
A few providers publish answers to these. Soniox offers EU data residency and says it does not retain audio or transcripts unless you ask it to. AssemblyAI offers an EU region, with retention and training settings to confirm. The OpenAI API can run in a Europe region with Zero Data Retention for eligible accounts, though abuse-monitoring logs may hold content for up to 30 days by default. Deepgram can be self-hosted on infrastructure you control, which suits teams more than solo coaches. Speechmatics is capable, but its terms need a careful read. Sonix is straightforward, though it stores data on US infrastructure. For most solo coaches, a local-first tool is the shorter path.
The rule I come back to
Use a transcription tool you trust, keep the audio on your own device where you can, remove the identifying details yourself, then analyse. That last step is where CoachFlow comes in, on a transcript you have already cleaned and chosen to use. It is the same workflow I describe in how to use AI for coaching sessions safely.
If you want to see what a cleaned transcript turns into, start a 7-day free trial at coachflow.space. You are not charged when you sign up, and you can cancel during the trial at no cost.