How LifeEcho Uses OpenAI Responsibly (And Why It Matters for Family Voices)

Most software companies are not transparent about what they send to AI vendors. They write a generic line in the privacy policy — "we use third-party AI services" — and move on.

That isn't good enough for a service that handles the recorded voice of someone you love. Voice memory is an intimate category. The content is often emotional, sometimes sacred. Families deserve to know exactly what we're doing with the recordings they trust us to preserve.

This is that document. It's written specifically for people who want the real answer, not the marketing answer.

The short version

LifeEcho uses two OpenAI services on every recording:

1. Whisper (OpenAI's speech-to-text model) to transcribe the audio into searchable text.
2. GPT (OpenAI's text model, currently GPT (currently gpt-4o-mini, configurable) by default) to generate a warm first-person title and a 1–3 sentence summary from the transcript.

Everything else — storage, user accounts, family sharing, prompts, pricing, search — runs on LifeEcho's own infrastructure, not OpenAI's. AI is a tool we use for two specific jobs. It is not where your recordings live.

What gets sent to OpenAI (exactly)

For transcription: The audio file itself, submitted to OpenAI's /v1/audio/transcriptions endpoint. We request the verbose_json response format with word-level timestamp granularity. Long recordings get split at silence boundaries on our side first, so each chunk sent to OpenAI is within their 25MB file limit. We do not attach any identifying user metadata to these requests.

For title and summary generation: The transcript text (not the audio — just the text), submitted to OpenAI's /v1/chat/completions endpoint with a structured prompt instructing the model to generate a warm, first-person title and a 1–3 sentence summary in first-person perspective. Again, no user identifying information is attached.

What gets sent with every API call: Our own API key for authentication, and the minimum required HTTP headers. That's it.

What does NOT get sent to OpenAI

• Your name, email, or account information. OpenAI does not know which user any particular recording belongs to.
• Your family members' names or contact information.
• Your payment or billing data (that lives entirely in Stripe).
• Other recordings in your library. Each API call is self-contained — one recording at a time.
• Metadata about who you share recordings with.
• Login credentials, passwords, or authentication tokens.

The OpenAI API is designed to be stateless for our use case. We send a specific piece of content, we get back a specific piece of output, and the call ends.

What OpenAI does (and doesn't) do with the data

OpenAI's API data policy — which is legally binding to them — states:

• API inputs and outputs are not used to train OpenAI's public models unless the customer explicitly opts in. LifeEcho does not opt in. The recordings and transcripts you submit are not used to improve GPT or Whisper for anyone else.
• Inputs and outputs are retained briefly (typically up to 30 days) for abuse monitoring, then deleted from OpenAI's systems. After that retention period, OpenAI does not have a copy of your content.
• Zero Data Retention agreements are available for certain enterprise customers; we're evaluating this for LifeEcho but have not needed it to date because the default API retention is already short.

These are the terms under OpenAI's public API. They are distinct from the consumer ChatGPT product, which has different defaults and does sometimes retain conversations for training. We deliberately use the API, not the consumer product.

Why we chose OpenAI (and what we'd switch to)

We evaluated speech-to-text and text-generation providers in roughly the following order:

For transcription, we compared OpenAI Whisper against Google Cloud Speech-to-Text, AWS Transcribe, Azure Speech, Deepgram, and AssemblyAI. Whisper came out ahead on:

• Accuracy with elderly voices and varied accents (the most important criterion for our use case).
• Handling of long-form single-speaker content (family stories, not meetings).
• Word-level timestamp quality.
• Simple, predictable pricing.

Deepgram and AssemblyAI were competitive on accuracy and would likely work well as backups if Whisper became unavailable. Google, AWS, and Azure were all slightly less accurate on the specific voice profiles that matter most to families.

For title and summary generation, we compared OpenAI GPT against Anthropic Claude and open-source models (Llama, Mistral). GPT and Claude both produced excellent first-person summaries with careful prompting. GPT was marginally better at the specific tone we wanted ("warm, first-person, feels like a scrapbook caption, not a meeting summary") and simpler to integrate. We re-evaluate this choice periodically. Claude is a likely secondary option if we expand provider diversification.

We explicitly rejected:

• Voice-cloning services (ElevenLabs, Play.ht, etc.). Not because they're bad products, but because we're not in the voice-cloning business. More on this in our post on why we don't do voice cloning yet.
• Consumer AI products (ChatGPT web, Gemini web, etc.) — these have different data policies than the APIs.

The specific policy choices behind our integration

Choice 1: API only, never consumer products. Consumer AI products (ChatGPT, Gemini web UI) may retain your inputs for training in ways that API calls do not. We exclusively use enterprise APIs.

Choice 2: Audio only to Whisper; transcript only to GPT. GPT does not need the audio — it needs the text. Sending the audio would be unnecessary additional data transmission. We send the minimum required for each task.

Choice 3: No user identifiers in API calls. We do not tag API calls with user IDs, emails, or other identifying information. Even if OpenAI's systems retained our calls longer than their policy states, those retained calls could not be tied back to specific LifeEcho users.

Choice 4: Calls are opt-out, not opt-in, but transparent. Every LifeEcho user's recordings are transcribed and summarized automatically — it's a core product feature, not an add-on. We're transparent about this (Privacy Policy, Terms of Service, and this article). If a family specifically requests that their recordings not be processed by AI, we can honor that request manually, but it's not our default workflow.

Choice 5: We own the storage; OpenAI doesn't. Your recordings, transcripts, titles, summaries, and all associated metadata live in LifeEcho's own infrastructure (Azure Blob Storage and Azure Tables). OpenAI holds the data briefly during processing and then discards it. The persistent copy is with us.

Choice 6: Model choice is configurable. We use GPT (currently gpt-4o-mini, configurable) for title/summary generation by default because it's fast, accurate enough for this task, and cost-effective. More capable (and expensive) models can be enabled for specific use cases without changing our data policies.

What this means for families

You own your recordings. They live in LifeEcho's storage, accessible to you through our dashboard, downloadable to your own device at any time. OpenAI briefly processes them and moves on.

Your loved one's voice is not being used to train AI models that will serve other customers. Even if Whisper gets better over time, it will not get better because of your grandmother's voice recordings.

You can download and leave. If you ever want to take your family's recordings elsewhere, you can download all of them, in their original audio format plus transcripts. We don't lock you in.

We're specific about what we do, and we'll stay specific. If our data policies change, we'll update this document. If we add new AI providers or features, we'll explain what and why. If we ever enter the voice-cloning business, you'll hear about it explicitly, not by discovering it in a feature update.

What responsible AI looks like in memory preservation

Most AI companies are still figuring out what "responsible" means in their category. For a voice memory service, our operating principles are:

1. The real voice is the asset. AI exists to make the real voice more useful, not to replace it. Transcription, titles, summaries, search, memoir export — all of these serve the authentic recording.
2. Minimum data, minimum time. Send only what's needed for the task, keep it with the vendor only as long as needed, discard aggressively.
3. Be specific, not generic. "We use third-party AI services" is not a meaningful disclosure. Naming the vendors, the models, the data flows, and the policies is.
4. Owe families the truth. Families are trusting us with the recorded voice of someone they love. Vague reassurance is worse than no reassurance. Specific reassurance is the only kind that's actually reassurance.
5. Choose providers who match our values. We use OpenAI because their API policies align with our needs today. If they stop aligning, we'll switch.

This isn't a permanent philosophy — AI is changing fast, and we'll update this document as new capabilities and new vendors emerge. But the underlying commitment doesn't change: your voice, your family's voice, the voices of the people you love — those are yours. We are a tool that helps you keep them. We are not the new owner of them.

Related: How AI transcription works for family phone calls · AI voice cloning vs real recordings · AI at LifeEcho