The short answer
To add on-device speech-to-text to an app, pick a runtime for your platform first and a model second: on iOS and macOS, Apple's SpeechAnalyzer (built in, OS 26 or later) or a Core ML model such as WhisperKit, FluidAudio's Parakeet or Phonon-2 Core ML; on Android and across platforms, whisper.cpp, sherpa-onnx, Moonshine or Vosk; on desktop and servers, a Python or C runtime such as the fermion package. Budget for a model of roughly 50 MB to 3 GB inside the app.
Smaller models fit more devices. Phonon-2.5 covers 25 European languages in 156 MB; Whisper base (74 M parameters, 142 MiB in whisper.cpp's format) covers 99 languages but has a much higher error on the same 25 FLEURS sets (53.71 % against 11.71 %).
Runtimes and models for app developers
Platforms, sizes and languages from each project's own pages, read 2026-11-13.
| Option | Platforms | Model size | Languages | Integration |
|---|---|---|---|---|
| Apple SpeechAnalyzer | iOS, iPadOS, macOS, visionOS 26+ | none in your app (system assets) | Apple's supported locales | Swift, Speech framework |
| WhisperKit | Apple silicon | Whisper tiny to large-v3 turbo (626 MB build recommended) | 99 (Whisper) | Swift package |
| FluidAudio | macOS, iOS | Parakeet TDT v3 or v2 | 25 European and Japanese (v3); English (v2) | Swift package, CocoaPods |
| Phonon-2 Core ML | Apple silicon Macs | Phonon-2 (164 MB model) | English | Swift and Python runners |
| fermion package (Phonon) | macOS, Linux x86-64 and arm64, Windows x86-64 | 156 to 164 MB | 25 European (Phonon-2.5) or English | Python, CLI, local HTTP server |
| Phonon Android package | Android arm64 | Phonon model | see release notes | native library |
| whisper.cpp | macOS, iOS, Android, Linux, Windows, WebAssembly | 75 MiB (tiny) to 2.9 GiB (large) | 99 (Whisper) | C/C++, Swift, Java, JS and more |
| sherpa-onnx | Android, iOS, HarmonyOS, desktop, WebAssembly, Raspberry Pi | per model | per model (Whisper, Parakeet, Zipformer, SenseVoice, Moonshine) | 12 languages incl. Kotlin, Swift, Dart |
| Moonshine | iOS, Android, macOS, Linux, Windows, web, Raspberry Pi | from about 1 MB | 10 for speech-to-text | Python, JS/WASM, mobile |
| Vosk | Android, iOS, Raspberry Pi, desktop through Python | about 50 MB per language | 20+ | Python and others |
Choosing by platform
- iPhone and iPad: SpeechAnalyzer if your users run iOS 26, because it adds nothing to your download. For a model you control, WhisperKit (Whisper) or FluidAudio (Parakeet, with diarization and VAD) run on the Neural Engine.
- Mac: the same options, plus Phonon-2 Core ML for English on the Neural Engine (606 times realtime on an M5 MacBook Air, Phonon-2's published figure) and MLX-based tools for the GPU.
- Android: whisper.cpp, sherpa-onnx, Moonshine and Vosk all ship Android paths; Phonon has an Android arm64 package.
- Windows and Linux desktop: whisper.cpp, sherpa-onnx and the fermion package, which runs the Phonon models on plain CPUs (143 times realtime on eight AMD Zen 5 cores, Phonon-2's published figure).
Size, memory and battery
Download size sets who installs the app; memory at run time sets which devices can keep the model loaded. whisper.cpp's README lists about 273 MB of memory for Whisper tiny and about 3.9 GB for large, so the largest models are a poor fit for phones.
The Neural Engine on Apple devices runs a model with less power than the GPU, which matters for features that listen for long periods. On other platforms, a small model on the CPU is often the practical choice.
Privacy: what on-device does and does not mean
On-device recognition means the model does not need to send audio to a server to transcribe it. Whether audio leaves the device is still a decision your app makes: analytics, crash reports, sync and opt-in data sharing can all upload it. Tell users plainly what is sent and give them a switch.
Open weights also let you check accuracy yourself: run the model on public test sets or on your own recordings before you ship it.
Questions
- What is the smallest accurate speech-to-text model for mobile apps?
- Among models with published numbers, Phonon-2.5 English (158 MB) averages 4.82 % word error on the seven Open ASR Leaderboard English sets, and Phonon-2.5 (156 MB) covers 25 European languages. Moonshine offers models from about 1 MB with less published accuracy data.
- Can I use Apple's built-in speech recognition in my app?
- Yes. The Speech framework's SpeechAnalyzer and SpeechTranscriber are available on iOS, iPadOS and macOS 26 or later, and use system-managed model assets rather than a model in your app bundle.
- What is the best way to run speech-to-text on Android?
- whisper.cpp, sherpa-onnx, Moonshine and Vosk all support Android, and Phonon has an Android arm64 package. Choose by languages and download size.
- Does on-device speech recognition work without internet?
- The recognition itself runs on the device once the model is installed. Whether your app needs a connection for anything else, such as sign-in or sync, is up to the app.
- How much memory does on-device speech recognition need?
- It depends on the model: whisper.cpp lists about 273 MB for Whisper tiny and about 3.9 GB for Whisper large. Models of 150 to 200 MB, such as Phonon-2.5, sit near the low end.
- Is on-device speech recognition private?
- It can be: the model transcribes without a server. Your app's own uploads, analytics and data sharing still decide whether audio leaves the device.
Related
Sources and method
- Apple SpeechAnalyzer and SpeechTranscriber: developer.apple.com/documentation/speech, read 2026-11-13.
- WhisperKit: github.com/argmaxinc/WhisperKit; FluidAudio: github.com/FluidInference/FluidAudio; whisper.cpp (sizes and memory): github.com/ggml-org/whisper.cpp; read 2026-11-13.
- sherpa-onnx: github.com/k2-fsa/sherpa-onnx; Moonshine: github.com/moonshine-ai/moonshine; Vosk: alphacephei.com/vosk; read 2026-11-13.
- Phonon sizes, platforms and Phonon-2 speed figures: Phonon model pages and /docs/speech/. No Phonon-2.5 speed figures are published.
- Phonon English averages: Fermion Research measurement with the Open ASR Leaderboard's scoring code on the full seven test sets (Phonon-2.5 release post, 2026-10-22).
- FLEURS numbers: Fermion Research measurement on the full FLEURS test set of each language, one scoring code and normalisation for every model (Phonon-2.5 release post, 2026-10-22).
Canonical: https://www.fermionresearch.com/speech/on-device-speech-to-text-for-apps/