Opt Out of AI Training With Your Audio: Per-Vendor Guide (2026)

Most transcription services don't train on your audio by default. But "most" is not "all," and the ones that do rarely lead with that fact. This guide walks through what each major provider actually does, verified against current policy pages as of July 2026, with the opt-out path where one exists.

Why It Matters More Than You Think

When your audio enters a training pipeline, two things happen that are hard to undo:

1. The model learns statistical patterns from your content, including domain vocabulary, speaker styles, and sometimes sensitive subject matter.
2. Removing your contribution later is not possible. You can delete the original file; the model's weights still carry its influence.

For podcasts and meeting notes, this is a low concern. For client calls, medical conversations, internal strategy, or source-protected interviews, it is a real exposure. The question is not just "is my audio stored?" but "is my audio teaching something?"

See is AI transcription private for the broader picture beyond training.

Three Categories of Provider

Category 1: Structural non-training. The service uses a pretrained model in inference mode. Your audio is processed and discarded. There is no training pipeline the provider could insert your data into, regardless of policy.

Category 2: Opt-in training. Training is off by default. You can volunteer data, sometimes in exchange for discounts or accuracy improvements.

Category 3: Default-on training. Your audio goes into the training pipeline unless you actively opt out. These providers require the most attention.

Category 3 is the worst posture. Category 1 is the simplest to trust. Category 2 is fine if you verify the default is genuinely off.

Per-Provider Reality

Policies change. Verify against current terms before relying on this table. Dates checked: July 2026.

OpenAI Whisper API

API data is excluded from training by default. OpenAI's enterprise privacy page commits that data submitted through the API is not used to train models. This applies to all API customers including the Whisper endpoint. No opt-out toggle is required because the default is already off.

Note the distinction: this is the API. OpenAI's ChatGPT consumer product has a different policy and has historically used conversation data for training unless opted out via account settings.

Zero Data Retention (ZDR) is available for eligible enterprise customers on supported endpoints, where data is processed in memory and not retained after the request. ZDR is not a self-serve toggle; it requires working with your account team.

Checked against: openai.com/policies, July 2026.

Deepgram

Training is opt-in through the Model Improvement Partnership Program. Standard customers are not enrolled by default. To explicitly exclude a request, add mip_opt_out=true as a query parameter to any API call. Data from opted-out requests is retained only long enough to complete the request.

Program participants receive discounted pricing; opting out does not incur a penalty, it simply means forgoing the participant discount. If you want no training risk at all, add the parameter and accept standard pricing.

Checked against: developers.deepgram.com/docs/the-deepgram-model-improvement-partnership-program, July 2026.

For a full pricing breakdown of Deepgram's tiers, see Deepgram Nova-3 explained.

Google Cloud Speech-to-Text

Data logging is opt-in. This is the opposite of what the previous version of this post implied. Customers who do NOT enable data logging have their audio processed and discarded; it is not stored for model improvement. Lower pricing is offered to customers who opt in.

If you have ever enabled data logging on a project, data logged before you disabled it stays in Google's systems. Future requests are not logged once you disable it, but historical logging cannot be recalled.

Path to verify: Cloud Console, APIs and Services, Cloud Speech API, Data logging tab. Confirm it is off. Capture a screenshot.

Checked against: docs.cloud.google.com/speech-to-text/docs/enable-data-logging, July 2026.

AWS Transcribe

AWS's Service Improvement Program is opt-in. Customer data is not used to improve models unless you actively enroll. Per AWS documentation, the default for API customers is no participation in service improvement training.

If your account has had any opt-in enabled, verify via the AWS Service Health Dashboard or the AWS console under your account preferences.

Checked against: AWS documentation, July 2026.

Otter.ai

Otter trains on de-identified customer audio by default. Its privacy policy explicitly states it uses "Meeting and Uploaded Information" including audio recordings and transcriptions to train its proprietary AI, after a de-identification process. Human reviewers are not involved, but the training itself is automatic and on by default.

My take: the de-identification claim is precisely what is being contested in an active federal class action (filed December 2025, ongoing as of July 2026) alleging Otter recorded private conversations and used them to train models without adequate consent. "De-identified" is not the same as "not used." Until that litigation resolves, I would treat Otter as Category 3 and assume training is on by default.

The opt-out path is not a simple settings toggle. You would need to contact Otter directly; the "Do Not Sell or Share My Personal Information" footer link covers California privacy rights broadly but does not cleanly map to audio training.

Checked against: otter.ai/privacy-policy, July 2026.

Rev.com

Rev uses customer data anonymously to train its own proprietary AI, per its privacy policy. This is not third-party LLM training; Rev states it does not share data with external AI providers for training. But its internal AI models do learn from customer transcripts.

Customers can opt out by emailing support@rev.com. There is no in-product toggle; you email the request and it goes into their records.

For human transcription at Rev, a different set of concerns applies: real people review your audio. Rev screens its transcriptionists and has them sign NDAs, but the point stands that human eyes (and ears) reach the content. Choose AI transcription at Rev if you want to avoid human review.

Opt-out path: email support@rev.com with a request to exclude your data from training.

Checked against: rev.com/security, support.rev.com/hc/en-us/articles/19509652584717-AI-Training-Opt-Out, July 2026.

Fireflies.ai

Fireflies explicitly prohibits AI training on customer meeting data. Its privacy policy (updated March 6, 2026) states it does not use personal information for AI model training and enforces zero data retention with its vendors, meaning third-party processors cannot store meeting content after delivery. Vendors are contractually prohibited from training on the data.

No opt-out is needed; this is a baseline policy, not an opt-in commitment.

Checked against: fireflies.ai/privacy-policy, July 2026.

Descript

Production models at Descript use no customer data. In-house R&D models use data only from users who have opted in to data sharing. There are no plans to use data from opted-out users at any stage.

The one nuance: AI voice features (the Overdub voice cloning tool) use voice data with de-identification for voice technology research. If you are using general transcription, this does not apply.

Checked against: descript.com/privacy, July 2026.

Happy Scribe

Happy Scribe does train machine learning models on customer content. Its privacy policy, under Section II (Content Data), states: "We will be able to store the Content for training our machine learning algorithms." The legal basis given is "legitimate interests" under GDPR.

No explicit opt-out mechanism is described in the policy document. Under GDPR, users can object to processing based on legitimate interests, but Happy Scribe has not published a self-serve path. Contact dataprotection@happyscribe.com if you need this in writing.

Happy Scribe is an EU-incorporated company with EU data centers, which does provide some procedural protection, but EU residency does not eliminate the training use.

Checked against: happyscribe.com/privacy, July 2026.

TurboScribe

TurboScribe's policy explicitly states it does not train AI or machine learning models on user media files or transcripts. No opt-out is needed. The service runs Whisper in inference mode.

Checked against: turboscribe.ai/privacy, July 2026.

How to Verify the Claim Is Real

A policy statement is a starting point, not a guarantee. The practical checks:

1. Read the current ToS and privacy policy. Look for the specific word "train" or "machine learning" in both documents. Vague language like "improve our services" without clarification is a signal to ask follow-up questions.
2. Check the privacy policy update date. A policy last updated in 2022 may not reflect current practices.
3. Request written confirmation. Email support and ask specifically: "Does my audio data enter any model training pipeline?" The response becomes a record.
4. Get a DPA with explicit non-training language. For business customers, the Data Processing Agreement is the contractual mechanism. Include language like: "Provider shall not use Customer Data, including audio recordings or transcripts, to train, fine-tune, or otherwise improve any artificial intelligence model, except with the express written consent of Customer." This is enforceable; a toggle is not.
5. Look at what model is used. If the provider lists Whisper Large-v3, a pretrained OpenAI model, the provider cannot easily fine-tune it on your data. Proprietary end-to-end models are harder to audit.

For more on what your contractual options look like in practice, see transcription and confidentiality agreements.

The Structural Non-Training Argument

The strongest privacy claim is not policy-based but architectural. If a service uses a foundation model like Whisper Large-v3 in pure inference mode, the provider lacks the infrastructure to incorporate your audio into training updates. Training Whisper requires the original OpenAI training setup; re-training on customer audio is not something a typical SaaS company does.

This is why pretrained-model providers like TurboScribe and tools built on the Whisper API sit in Category 1. The claim is structural, not just a statement in a document.

If you need maximum assurance, on-device vs cloud transcription covers the case for self-hosting Whisper on your own hardware, where the model never leaves your environment.

ConvertAudioToText's Position

If you need a clean transcript without a meeting bot or a proprietary training pipeline touching your data, ConvertAudioToText uses Whisper Large-v3 and AssemblyAI in inference mode only. Neither pipeline is retrained on customer audio. Business accounts can request a DPA with explicit non-training language.

The Decision Framework

1. Is your audio sensitive? If not, most reputable providers work fine.
2. If yes: check whether the provider defaults to training (Otter, Happy Scribe, Rev's proprietary AI) and either opt out or choose a different provider.
3. Do you need contractual non-training language? Get a DPA that includes it, not just an app toggle.
4. Do you need structural non-training assurance? Choose a provider using pretrained models in inference mode (Whisper-based tools, TurboScribe, Fireflies).
5. Is the content extraordinarily sensitive? Self-host Whisper on your own hardware.

For GDPR-regulated contexts, see GDPR-compliant transcription for the regulatory framing, which involves more than training policy.

FAQ

Does OpenAI use Whisper API audio to train its models?

No. OpenAI's enterprise privacy policy states that data submitted through the API is not used to train models. This applies by default to all API customers, including those using the Whisper endpoint. You do not need to do anything to opt out; the default is already off. Zero Data Retention is available for enterprise customers who want additional assurance.

How do I opt out of Deepgram's model training?

Deepgram's Model Improvement Partnership Program is opt-in, meaning you are not enrolled by default. If you want to explicitly exclude individual API requests, add mip_opt_out=true as a query parameter. Data from opted-out requests is retained only for the duration needed to process the request. Program participants receive discounted pricing, but opting out does not trigger any penalty.

Does Otter.ai train on my meeting recordings?

Yes, per Otter's privacy policy. Otter trains its proprietary AI on de-identified audio recordings and transcriptions from customer meetings. The training is automatic and default-on. An active federal class action filed in 2025 challenges whether this practice met informed consent standards. If training on your meeting content is unacceptable, this is a firm reason to choose a different provider.

Is Google Cloud Speech-to-Text data logging on by default?

No. Data logging for Google Cloud Speech-to-Text is opt-in. Unless you have actively enabled it in your Cloud Console, your audio is processed and discarded. Customers who opt in get lower pricing in exchange for contributing data. If you have an existing project where logging may have been enabled previously, verify the current setting under Cloud Console, APIs and Services, Cloud Speech API, Data logging tab.

What is the strongest protection against a vendor training on my audio?

A combination of two things: a DPA with explicit contractual language prohibiting training, and choosing a provider whose technical architecture makes training on your data structurally infeasible (such as one using a pretrained Whisper model in inference-only mode). A DPA is enforceable in court; an app toggle is not. Structural non-training (no training infrastructure, pretrained-only model) is harder to circumvent even if a policy later changes.

Sources

• OpenAI data usage policy: openai.com/policies/how-your-data-is-used-to-improve-model-performance
• Deepgram Model Improvement Partnership Program: developers.deepgram.com/docs/the-deepgram-model-improvement-partnership-program
• Google Cloud Speech-to-Text data logging: docs.cloud.google.com/speech-to-text/docs/enable-data-logging
• Google Cloud data logging terms: cloud.google.com/speech-to-text/docs/data-logging-terms
• Otter.ai privacy policy: otter.ai/privacy-policy
• Otter.ai litigation coverage: ktslaw.com/en/Blog/GlobalPrivacy-and-CybersecurityLaw/2025/9/Otterai-Suit-Highlights-Risks-of-Using-User-Data-to-Train-AI
• Rev.com AI training opt-out: support.rev.com/hc/en-us/articles/19509652584717-AI-Training-Opt-Out
• Fireflies.ai privacy policy (updated March 6, 2026): fireflies.ai/privacy-policy
• Descript privacy: descript.com/privacy
• Happy Scribe privacy policy: happyscribe.com/privacy
• TurboScribe privacy policy: turboscribe.ai/privacy