Transcription Apps That Don't Train on Your Data

Data checked: 2026-10-21

Whether a transcription service may train its models on your recordings is a policy question, not a feature, and the answer is written somewhere different at every vendor: a pricing page, a security page, a privacy policy or an FAQ. We record one of five states for every tool: “no” (the vendor states it does not train), “only if you opt in”, “yes, unless you opt out”, “yes” (trains, no opt-out described) and “unclear” (the pages we checked don’t say). Everything below quotes what the vendor itself publishes, checked in October 2026; we have not audited anyone’s servers.

Stated no-training, in policy or pricing pages

• Sonix: security page, “Sonix never uses the audio, video or transcripts you process to train its models, on any plan or trial”, and deleting a file wipes the audio and transcript.
• Fireflies.ai: pricing page, “Your data is never used for AI training”.
• tl;dv: privacy policy section 9 excludes recordings, transcripts and notes from training, including at its AI providers.
• Transkriptor: security page rules out training its own or third-party models.
• Good Tape: privacy policy says files are never used for training.
• jamie: privacy policy section 1.8 excludes meeting content and personal data from training or fine-tuning, including at providers.
• Trint: homepage states Trint never trains its algorithms on transcripts.
• Rimo Voice: every plan on the pricing page lists “no AI training” (AI学習なし).
• toruno: terms have excluded AI development and training use since 14 March 2024.
• Amberscript: privacy policy section 1.1: generated audio and transcripts are never used to train the service’s language models.
• Bluedot: privacy policy section 3.3: customer content is not used to train any AI models.
• TurboScribe: its FAQ says it never trains on your content and that you can delete your data; we have not yet been able to read the full privacy policy, so this one rests on a weaker page than the others.
• superwhisper: policy states data is not used for AI training, with third-party services excluded from that promise.

Training only if you opt in

These services keep your content out of training unless you actively agree, which is safe by default but worth knowing at signup:

• Happy Scribe asks at signup; if you opt in, de-identified content may also be licensed or shared with partners to train their models.
• PLAUD: “Plaud will not train on your data by default, explicit opt-in is required.”
• OpenAI Speech to Text API: API data is not used unless you explicitly opt in to share it.
• Google Cloud Speech-to-Text: audio and transcripts are not logged by default; opting into data logging permits training use.
• Otolio（旧：スマート書記） permits machine learning with the customer’s consent.
• Read AI runs its Customer Experience Program only for users who opt in.

Offline apps: nothing to train on

On-device apps never upload recordings for transcription: MacWhisper (“no data leaves your machine” with local models), Vibe, Buzz, Aiko, noScribe, aTrain and whisper.cpp. The trade-offs are hardware speed and fewer team features; see offline transcription software.

The other end of the scale

Otter.ai and Rev train by default: Otter’s policy describes no opt-out, and Rev’s agreement says content continuously trains its speech models (its help center says Enterprise and Unlimited customers can opt out). Tools that train unless you find and use an opt-out include Fathom, Descript, Granola, AssemblyAI, Deepgram and Amazon Transcribe. Use the “don’t train on your data by default” filter in the directory to slice the whole list.

How we record and source every privacy claim is described in how we check; each tool’s page links the exact vendor document behind its entry.