Glossary
Federated learning
Federated learning is a way to train one machine learning model on data spread across many devices. Each device computes an update from its own data, and only those updates are sent to a server, so the raw data does not have to be collected in one place.
Updated October 17, 20262 cited facts, each traced to its source
Also called on-device training.
Federated learning trains a shared model without collecting the training data in one place. Each device uses its own data to compute a change to the model, and only those changes are combined on a server. The term comes from McMahan and colleagues, who proposed in 2017 leaving the training data on the mobile devices and learning a shared model by aggregating locally computed updates 1.
How one round of training works
A federated system repeats the same round many times:
- The server sends the current version of the model to a set of devices.
- Each device trains the model briefly on the data it holds, such as the text a person typed.
- Each device sends back only the change to the model, not the data.
- The server averages the changes into a new model, and the cycle starts again.
The averaging step is the core of the method. The McMahan paper calls it iterative model averaging, and it reports experiments across five model architectures and four datasets 1.
Phone keyboards showed that the approach works
Phone keyboards are a useful test case, because the text people type is personal. In 2018, researchers showed that a language model can be trained on client devices without exporting sensitive user data to servers 2.
A concrete example: a phrase learned across many phones
Imagine that many people type "see you Thursday" after a meeting invitation, and the keyboard model should predict that phrase. Each phone uses its own typing to compute a small change that makes the phrase more likely. The server receives many of these changes, averages them, and sends back a better model. The server never receives the messages themselves. This is a hypothetical illustration, not a description of any product's data flow.
Why it matters for dictation
Federated learning is one way a dictation model can improve from real use without the audio or transcripts leaving the device. It does not happen automatically. A product has to use this method, and its privacy policy should say whether it sends audio, transcripts, or only model updates.
The updates are derived from your data, so the method reduces what leaves your device rather than removing exposure entirely. When you compare dictation apps, ask three questions: does the app train on your audio, does it send transcripts to train, and can you turn that off? The answers matter more than the name of the method.
Related terms
- On-device speech recognition: recognition that runs on the device, which is the setting where federated updates usually start.
- Special category data (GDPR): the category that health and identity content can fall into when it is used for training.
Guides that cover it
Evidence
Every number on this page, with where it comes from. Open a fact to see the exact words in its source.
- 1.McMahan and colleagues proposed federated learning, an approach that leaves the training data distributed on mobile devices and learns a shared model by aggregating locally computed updates.
H. Brendan McMahan et al., Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), JMLR Workshop and Conference Proceedings volume 54, 2017 · Original
- 2.
Related terms
- Speech to Text Privacy: Where Your Voice GoesThe topic this term belongs to
- All-party consentAll-party consent is a recording rule under which every participant in a conversation must agree before it can be recorded or intercepted. California, Florida and Maryland use it in their statutes, and it is stricter than one-party consent.
- Attorney-client privilegeAttorney-client privilege protects confidential communications between a lawyer and their client about seeking legal advice or services. A third party who is present, or a tool that handles the conversation, can put that protection at risk.
- Biometric dataBiometric data is personal data produced by technical processing of a person's physical, physiological or behavioral characteristics that allows or confirms who that person is. A voice recording becomes biometric data when it is processed to identify the speaker.
- Business associate agreement (BAA)A business associate agreement (BAA) is the written contract that HIPAA requires before a covered entity lets a vendor create, receive, maintain or transmit protected health information on its behalf. Dictation tools used in clinics often need one, and a BAA alone does not make a product compliant.
- False accept (accidental activation)A false accept, also called a false wake or accidental activation, is when a voice assistant or dictation feature starts listening or recording although the wake word was not said. Smart speakers have been measured to wake by accident, and that can capture speech the user did not mean to share.
- On-device speech recognitionOn-device speech recognition converts speech to text with a model that runs on your own computer or phone, so the audio does not need to be sent to a server. It is also called local, offline or on-device dictation.
- One-party consentOne-party consent is a recording rule under which a conversation can be recorded lawfully when one of its participants agrees, which is usually the person doing the recording. The US federal wiretap law uses this rule, and individual states may set stricter ones.
- Protected health information (PHI)Protected health information (PHI) is individually identifiable health information that HIPAA protects when it is held or sent in electronic or other form by a covered entity or its business associate. Patient names with diagnoses, in dictated notes or voice recordings, are PHI. It is also called ePHI when held electronically.
- Speaker verificationSpeaker verification is a voice biometric check in which a system accepts or rejects a claimed identity by comparing a voice with stored models of the person claimed. It is also called voice authentication, voice ID or voice biometrics.
- Special category data (GDPR)Special category data is the group of personal data that Article 9 of the GDPR bans organizations from processing unless a listed exception applies. It includes health data and biometric data used to identify a person, and a voice recording can fall into it when it is processed that way.
- Voice cloningVoice cloning is the use of a model trained on recordings of one person to generate new speech in that person's voice. Some products need only minutes of clear audio, and cloned voices have been used in fraud, so recordings of your voice carry a real risk.
- VoiceprintA voiceprint is a stored model of a person's voice, built from measurable features of their speech and used to recognize or verify that person. Illinois law lists voiceprints among biometric identifiers, and other vendors call the same idea a voice template or speaker model.