Do cloud speech-to-text APIs store your audio?

Cloud speech-to-text services do not share one answer. The same vendor can keep nothing in one mode and keep files in another, and a default in one product can be an opt-in in another. The safe approach is to read the mode and the logging setting for the exact service you call, then set a retention rule you control, rather than assume that a speech API forgets your audio after it returns text.

Azure's real-time service keeps no audio at rest, but its batch service leaves storage to you

Microsoft documents that for real-time speech to text in Azure, audio is processed only in server memory, and no data is stored at rest ¹. That statement depends on two conditions: endpoint logging is off and no customizations are used ². The statement is about Azure AI Speech, not about dictation in Windows or Microsoft 365, which are covered in the question on Microsoft and voice recordings.

Batch transcription is different. Microsoft says the customer chooses where the audio and transcripts are stored and controls their retention ³. To set how long the generated transcript files last, you use a parameter called timeToLive ⁴. The page names that parameter for the generated text files. It does not give a default retention period for the input audio, so set your own storage rule for the audio.

Microsoft also notes that audio of people speaking, and its transcripts, may be considered personal data or sensitive data under various privacy regulations and laws ⁵. That is why retention matters even when a vendor offers a fast, accurate service.

Google Cloud Speech-to-Text logs nothing by default, and logged data has human access

Google says Cloud Speech-to-Text does not log customer audio or transcripts unless the customer opts in to data logging ⁶. A project that never enables data logging keeps no logged audio under that program.

If you do opt in, the access rules change. Google says only a limited set of authorized Google employees and contractors can access customer data collected under Speech-to-Text data logging ⁷. Access is limited, not zero. The page does not say how long logged data is kept, so ask before you enable logging.

Amazon Transcribe has job data and artifacts that you can delete

Amazon says customers can use its Delete APIs to delete the data and other artifacts associated with transcription jobs ⁸. The FAQ does not say how long Amazon keeps that data if nobody deletes it, so do not assume a time limit. Plan to delete on your own schedule.

An organization can also opt out of having its Transcribe content used to improve Amazon Transcribe and its other machine-learning services, by using an AWS Organizations opt-out policy ⁹. This controls training use, and it is a separate choice from deleting. A single account outside an organization needs its own settings checked, because the policy is organization-wide.

A table makes the three services easier to compare

The table is a summary, not a substitute for each vendor's current terms. Services change their defaults and product names, so check the page for the exact service you call.

Questions to answer before you send audio to any cloud API

1. Which mode is it? Real-time and batch services can store different things. Check which one your code calls.
2. Is logging or improvement use on? Look for the data-logging setting on Google Cloud, or the AWS Organizations opt-out for Transcribe.
3. How long is data kept, and how is it deleted? Look for a timeToLive setting on Azure batch and for Delete APIs on AWS.
4. Who can access the data? Google's logged data is accessible to a limited set of its own staff and contractors, so decide whether that is acceptable.
5. Do you need a contract? For health, legal or regulated content, check whether the vendor will sign the agreement your rules require, and who decides that you need one. Our guide to HIPAA-compliant dictation explains what to ask a vendor.

If you need the audio to stay on the device, a cloud API is the wrong tool, because any cloud service sends the audio to a server before it returns text. On-device recognition avoids that step. Yap, which we make, is an on-device dictation app for Mac and Windows, and you can see the download options. For how the recognizer itself works, see automatic speech recognition.

Where to go next

If you are choosing a service, read the question on whether OpenAI stores Whisper audio and the question on the fastest speech-to-text API. If you are checking a consumer product instead, read the question on Google and voice recordings, the question on Microsoft and voice recordings and the question on Alexa. The full picture is in our guide to voice data retention, part of our overview of speech to text privacy.