Voice Assistant Privacy: Where Your Data Goes

October 6, 2026 · privacy guide · 10 minutes read

Every time you say "Hey Google" or "Alexa", your voice travels through the internet, gets transcribed, analyzed, and stored. The companies say it's for "improving the service." But where exactly does your voice go? Who can access it? And can you opt out?

The Three Levels of Voice Data

Voice assistants collect data at three levels:

The recording itself: Raw audio of your voice
The transcription: Text version of what you said
Metadata: When you spoke, what device, your location, what action was taken

Different assistants handle these differently. Let's break it down.

Google Assistant / Gemini

What Gets Collected

Audio recordings: Stored by default, linked to your Google account
Transcriptions: Stored indefinitely
Context: Device ID, location, time, search history, app usage
Action results: What you asked for and what Assistant did

Who Has Access

Google's AI systems: Used to train models and improve recognition
Google employees: Can access recordings for "quality control"
Third-party contractors: Google admits some recordings are reviewed by humans
Law enforcement: Can be subpoenaed (happened in murder investigations)

How Long It's Stored

By default: forever, until you manually delete it.

You can set auto-delete after 3, 18, or 36 months, but it's opt-in (Settings → Google Account → Data & Privacy → Web & App Activity).

Can You Opt Out?

Partially:

You can disable "Web & App Activity" — but then Assistant loses context and becomes less useful
You can delete recordings manually via myactivity.google.com
You can't use Assistant without a Google account

Amazon Alexa

What Gets Collected

Audio recordings: Stored by default
Transcriptions: Stored indefinitely
Purchase history: What you bought via Alexa
Smart home logs: Every command to your devices

Who Has Access

Amazon's AI systems: Training data for Alexa improvements
Human reviewers: Amazon admitted in 2019 that thousands of employees listen to recordings
Amazon marketing: Used for product recommendations
Law enforcement: Subpoenable (Amazon has fought some requests, but complied with others)

How Long It's Stored

By default: forever.

You can enable auto-delete (3 or 18 months), but it's hidden deep in settings.

Can You Opt Out?

Partially:

You can disable "Help improve Amazon services" — but Alexa still stores recordings
You can delete recordings in the Alexa app → Settings → Alexa Privacy
You can't use Alexa without an Amazon account

Apple Siri

What Gets Collected

Audio recordings: NOT stored by default (Apple's main privacy selling point)
Transcriptions: Stored with a random identifier, not linked to Apple ID
Metadata: Device type, approximate location, interaction logs

Who Has Access

Apple's AI systems: Training data (but anonymized)
Apple employees: Can review transcriptions, but not recordings (unless you opt in)
Law enforcement: Apple can't provide recordings (they don't have them), but can provide transcriptions if subpoenaed

How Long It's Stored

Transcriptions: up to 6 months, then anonymized further.

Audio (if you opt in): 2 years.

Can You Opt Out?

Yes:

Siri doesn't store recordings by default
You can disable "Improve Siri & Dictation" to prevent transcription analysis
You can use Siri without iCloud (though some features break)

Amalia (Open Source)

What Gets Collected

Audio recordings: NOT stored at all — audio is sent to Groq for STT, then discarded
Transcriptions: Stored locally on device (encrypted DataStore)
Conversation history: Stays on your phone, never leaves unless you manually export
Metadata: None. No analytics, no telemetry, no crash reporting

Who Has Access

You: Only you. Conversations are encrypted and stored locally
Groq (during STT): Receives audio for transcription, processes it, returns text — no storage per their privacy policy
Groq (during LLM): Receives transcription + system prompt, generates response — processed in memory, not stored
Fish Audio (during TTS): Receives text, synthesizes speech, streams audio — no storage

How Long It's Stored

On device: Until you clear app data
On Groq/Fish Audio servers: Zero. Requests are processed in memory and discarded

Can You Opt Out?

Fully:

No account required
No analytics to opt out of (there are none)
Source code is open — you can verify nothing is collected: github.com/kurumi-mProject/amalia

Privacy Comparison Table

Feature	Google	Alexa	Siri	Amalia
Stores audio recordings	✓ (default)	✓ (default)	✗ (opt-in)	✗
Stores transcriptions	✓ Forever	✓ Forever	✓ 6 months	Device only
Human reviewers listen	✓	✓	✗ (no audio)	✗
Linked to account	✓ Google ID	✓ Amazon ID	Random ID	No account
Used for ads/marketing	✓	✓	✗	✗
Can be subpoenaed	✓	✓	Partial	Device only
Open source (auditable)	✗	✗	✗	✓ (MIT)

Real-World Privacy Incidents

2018: Amazon Alexa sent private conversation to random contact

A Portland couple's conversation was recorded, transcribed, and sent to one of the husband's employees without their knowledge. Amazon confirmed it was a "rare" bug.

2019: Google contractors listen to Assistant recordings

Belgian news outlet VRT revealed that Google contractors listen to recordings to improve speech recognition. Google suspended the program, then resumed it with an opt-out option.

2021: Siri recordings reviewed without consent

Apple paid $95M settlement after admitting contractors listened to Siri recordings that included private medical information, drug deals, and sexual encounters.

2022: Police subpoena Alexa recordings in murder case

Arkansas court ordered Amazon to hand over Alexa recordings as evidence. Amazon initially refused, then complied.

The Open Source Difference

With Amalia, privacy is verifiable:

Source code is public: github.com/kurumi-mProject/amalia
No analytics SDK, no crash reporting, no telemetry
Conversations stored in encrypted DataStore — you can inspect the database
API keys are yours (Groq + Fish Audio) — you control access logs

You don't have to trust the developer's word — you can audit the code yourself.

How to Protect Your Privacy

If You Use Google/Alexa/Siri:

Delete your recordings regularly: Google: myactivity.google.com | Alexa: App → Privacy | Siri: Settings → Siri & Search → Delete History
Disable human review: Opt out of "improve service" programs (buried in settings)
Use auto-delete: Set shortest retention period available
Mute when not in use: Physical mute buttons exist for a reason

If You Want Real Privacy:

Use open source alternatives: Amalia (MIT license), Dicio (GPL), Rhasspy (offline)
Self-host everything: Whisper.cpp + local LLM + Piper TTS = fully offline (guide: offline voice assistants)
Use your own API keys: With Amalia, you control Groq/Fish Audio keys — check their access logs yourself

The Bottom Line

Every mainstream voice assistant collects data. The difference is:

Google/Alexa: Collect everything, store forever, use for ads, human reviewers listen
Siri: Doesn't store audio by default, but still collects transcriptions
Amalia: On-device storage only, no account, open source, auditable

If privacy matters to you, use open source. If you're okay with the trade-off, at least configure auto-delete and disable human review.

Want to try a privacy-first assistant? Download Amalia and see what it's like when your conversations stay yours.