How On-Device Transcription Works on iPhone

Research conducted October 2026 from official documentation, research papers and legal references. Details may change.

Short answer: on a modern iPhone, a speech model stored on the phone turns audio into text without sending the recording to a server. Apple's newer SpeechAnalyzer API, introduced at WWDC25, is built for long-form audio like meetings and lectures ^[1][2].

What "on-device" actually means

Speech-to-text has two halves: capturing sound and running a model that maps sound to words. In cloud apps, the audio file is uploaded and a server runs the model. With on-device transcription, the model file sits on your iPhone and the neural engine and CPU do the work locally. Apple's older SFSpeechRecognizer API let developers force this with the requiresOnDeviceRecognition flag ^[4]. The newer SpeechAnalyzer framework, with its SpeechTranscriber module, is designed to run on device and handle long, conversational audio ^[1][3].

The practical results: it works in airplane mode, there's no upload wait, and the raw audio never leaves the phone.

The pipeline, step by step

1. Capture. The microphone records audio, usually at 16 to 48 kHz. Apps can turn on Apple's voice processing on the audio input, which applies echo cancellation and other cleanup tuned for speech ^[10].
2. Clean-up. Noise suppression reduces steady background sounds like fans and traffic. Research systems like RNNoise show how a small neural network can do this in real time on modest hardware ^[9].
3. Voice activity detection (VAD). A lightweight model flags which stretches contain speech, so the recognizer skips silence. Silero VAD is a widely used open-source example ^[8].
4. Recognition. An acoustic and language model turns speech into words. Modern models are trained end to end on huge amounts of audio. OpenAI's Whisper, for example, was trained on 680,000 hours of multilingual audio ^[5].
5. Formatting. The model, or a step after it, adds punctuation and capitalization and produces timestamps so text can be linked back to the audio ^[2][3].

Apple SpeechAnalyzer vs. Whisper

Both are strong. Apple's API saves each app from shipping its own large model file, and the system keeps it updated. Whisper gives developers full control and covers many languages ^[5][6].

What on-device transcription can't do (yet)

• Who said what. Labeling speakers ("speaker diarization") is a separate problem from recognition. See our guide Speaker diarization explained.
• Heavy crosstalk. When people talk over each other, accuracy drops for every recognizer.
• Rare names and jargon. Expect to fix product names and acronyms.

How Keepsay uses it

Keepsay supports 12 recording languages and transcribes on your iPhone with Apple's speech model, with a one-time model download per language and no account for on-device use ^[11]. If you choose Cloud AI for summaries, only the transcript text is sent, never the audio ^[12]. Keepsay also shows live transcription with a translation side by side.

Tips for better on-device transcripts

• Put the phone in the middle of the table, screen up, away from laptops and fans.
• Use one language per recording when you can.
• Keep the recording going through pauses; VAD will skip the silence ^[8].

Related: iPhone meeting recorder apps in 2026

Sources & methodology