Language Models for iOS 18 Call Translation

Apple's iOS 18 introduces real-time call recording, audio transcription, and multilingual processing directly within the Phone and Notes apps. The system converts spoken audio into written text and translates or summarizes that content into secondary languages using a pipeline of Apple’s proprietary foundation models. These capabilities rely on a combination of compact, on-device language models and larger server-based models under the Apple Intelligence framework.

The Core Models: Apple Foundation Models (AFM)

The primary engine translating and processing call transcriptions in iOS 18 is the Apple Foundation Model (AFM) family. Rather than relying on third-party solutions such as OpenAI's GPT models or Google's Gemini, Apple deploys its own specialized architecture:

  • AFM-on-device: A compact language model with approximately 3 billion parameters. It is quantized and optimized to run locally on the Apple Neural Engine (ANE) in devices powered by the A17 Pro chip or M-series processors. This model performs real-time language translation, contextual understanding, and text generation entirely on the device.
  • AFM-server: A larger cloud-based foundation model deployed via Private Cloud Compute (PCC). When a secondary language translation or detailed summarization task requires complex contextual processing beyond the capacity of local hardware, the transcript is securely routed to AFM-server running on dedicated Apple Silicon servers.

The Speech-to-Text and Translation Pipeline

Translating live call transcriptions involves two distinct machine learning steps:

  1. Acoustic Transcription: Before translation can occur, incoming audio is processed by Apple’s on-device speech recognition models. These models use Conformer-style neural network architectures optimized for the Neural Engine to convert raw audio into high-accuracy text streams in real time.
  2. Neural Machine Translation (NMT): Once the speech-to-text model generates the primary language transcript, the text tokens are fed into Apple's on-device Neural Machine Translation system. This translation architecture works in tandem with AFM-on-device to parse linguistic structure, retain conversational nuance, and generate the target secondary language.

Task-Specific LoRA Adapters

To make the on-device model capable of translation and summarization without consuming excessive memory, Apple utilizes Low-Rank Adaptation (LoRA) adapters.

The base 3-billion-parameter model remains frozen in device memory, while small, modular adapter layers (typically tens of megabytes in size) are swapped dynamically. When a user requests a translation or summary of a call transcript, the system activates the specific translation and summarization adapter weights. This architecture ensures high-speed, localized translation with minimal battery consumption and strict data privacy.