Does Azure speech to text store your data?

For real-time speech to text, the answer is no when endpoint logging is off and you use no customizations ¹. In that case Microsoft says no data is stored at rest ². Batch transcription and custom speech are different, because they store data in storage you choose or in Microsoft's storage. Which answer applies depends on the Azure feature your app calls.

Real-time speech to text keeps nothing at rest when logging is off

"At rest" means data kept on disk or in a database after the request ends. For real-time speech to text, Microsoft says audio is processed only in server memory, and no data is stored at rest ².

Microsoft also says it does not use customer data to improve its Speech models ³. That is a statement about training. It is separate from how long data is kept, so check both.

The no-storage statement has a condition. Microsoft says that if endpoint logging is disabled and no customizations are used, then no customer data is stored ¹. Endpoint logging and customizations, such as custom models, are the two settings that can change the answer. If either is on, do not assume the no-storage statement applies.

Batch transcription stores data in storage you choose

Batch transcription works differently. Microsoft says the customer controls where the data is stored, including how long it is kept ⁴. If you point the output at your own storage, the files stay there under your rules.

If the output goes to Microsoft's storage, you have two ways to delete it. You can call a deletion API, or set the timeToLive parameter so the data is deleted automatically after a time you set ⁵.

Custom speech stores your uploaded data by default

Custom speech and custom voice are the case most people miss. Microsoft says that by default your data is stored in Microsoft's storage, and your Speech resource uses Microsoft-managed encryption keys ⁶. You can prepare your own storage account for this data instead, which gives you control over where it sits.

Which Azure feature you use decides the answer

What to do

1. Identify the feature. Find out whether your code calls the real-time service, the batch API or a custom model. Each has different rules.
2. Keep logging off unless you need it. For real-time use, leave endpoint logging off, and avoid customizations you do not need.
3. Set a limit for batch output. Write output to your own storage with a retention rule, or set timeToLive on each job. You can also call the deletion API once you have the transcripts.
4. Decide where training data lives before you upload it. Custom speech data goes to Microsoft's storage unless you prepare your own.
5. Check regulated content first. If your speech includes health, legal or other regulated information, read Microsoft's terms with your compliance team before you send any audio.

Consumer Windows voice typing is a separate product with its own rules. See whether Microsoft keeps your voice recordings. For the same question about other cloud APIs, read whether speech to text APIs store audio. Our guide to how long each company keeps voice data compares Apple, Google, Microsoft and Amazon, and our overview of speech to text privacy covers where your voice goes in general. A setup that stores nothing is called zero data retention.