Live, on-device streaming speech-to-text on iOS, built on
parakeet.cpp. Tap the mic, speak, and the
transcript streams onto the screen — waveform on top, mic button in the middle,
transcript at the bottom, with a one-tap copy button per pane. Includes a
compare mode that runs Parakeet next to Apple's SpeechTranscriber (iOS 26)
and Moonshine side by side.
Community example — not maintained or tested by the parakeet.cpp core team (they have no Apple toolchain). Best-effort, maintained here.
parakeet.cpp is a pinned git submodule at third_party/parakeet.cpp.
scripts/build_xcframework.sh compiles libparakeet (+ ggml, Metal) from it into a
static Parakeet.xcframework for device + simulator. The app drives the streaming
C-API (parakeet_capi_stream_*) over mic audio captured with AVAudioEngine and
resampled to 16 kHz.
- Full Xcode (not just Command Line Tools) — needed for the iOS SDK and the
Metal toolchain (run
xcodebuild -downloadComponent MetalToolchainif the metal compiler is missing). - XcodeGen (
brew install xcodegen). - An iPhone (A16+ recommended for the 0.6B model in real time) and an Apple ID for signing.
git clone https://github.com/Kashif-E/parakeet-ios-demo
cd parakeet-ios-demo
./setup.sh # downloads prebuilt framework + models, generates the project
open ParakeetDemo.xcodeproj # set your signing Team, pick your iPhone, Runsetup.sh fetches the prebuilt Parakeet.xcframework (from the latest release),
the Parakeet model (~685 MB) and Moonshine base-en, then runs XcodeGen — no
submodule, no Metal toolchain, no cross-compile. All you add is your Apple
signing Team in Xcode (required for any on-device iOS app).
To build the framework yourself instead of downloading it:
git clone --recursive https://github.com/Kashif-E/parakeet-ios-demo # --recursive for the submodule
cd parakeet-ios-demo
./scripts/build_xcframework.sh # needs full Xcode; -> vendor/Parakeet.xcframework
./scripts/generate.sh
# then add models per "Models" below, open, Run.GGUFs are on Hugging Face — kashif3314/nemotron-3.5-asr-streaming-0.6b-gguf (unofficial, converted with parakeet.cpp; weights are OpenMDW-1.1):
| file | recipe | size | stock parakeet.cpp? |
|---|---|---|---|
…-q4_k.gguf |
q4_k linear, F32 rest | ~685 MB | ✅ use this |
…-q4k-f16.gguf |
q4_k linear, f16 rest | ~486 MB | |
…-q4k-q8.gguf |
q4_k linear, q8_0 rest | ~460 MB |
Download the q4_k file and save it as ParakeetDemo/Resources/model.gguf — it
loads on the vanilla pinned submodule, no patch needed.
For compare mode, also fetch Moonshine base-en (encoder_model.ort,
decoder_model_merged.ort, tokenizer.bin) from
moonshine-ai/moonshine into
ParakeetDemo/Resources/base-en/.
The q4k-f16 / q4k-q8 files store the non-matmul tensors (conv/LSTM/featurizer/
norm) in reduced precision (~30 % smaller, WER 0 vs NeMo on tested clips) and
rely on a load-time dequant that isn't in released parakeet.cpp yet. To use them,
apply parakeet-rest-loader.patch (also in the HF repo) to the submodule before
building:
cd third_party/parakeet.cpp
git apply /path/to/parakeet-rest-loader.patch # adds: quantize --rest f16|q8_0 + load-time dequant
cd ../.. && ./scripts/build_xcframework.shThe patch modifies MIT-licensed parakeet.cpp and is provided under the same MIT license.
Toggle "Compare" to run three engines on the same mic audio, stacked:
Parakeet (this) · Apple SpeechTranscriber (iOS 26, system) · Moonshine
(SPM package, bundled base-en model). Each pane has a copy button. Useful for
judging accuracy, punctuation, number formatting, and latency for your use case.
- Uses a stock parakeet.cpp model so it builds from the vanilla pinned submodule
with no extra patches. (A smaller
q4_k + q8-restbuild exists but needs a load-time dequant patch; not used here.) - The framework embeds the Metal shaders (
EMBED=ON) so it's self-contained — no Metal toolchain or submodule needed to use it. Shaders compile once at first launch (a few seconds, behind the loading state). - The model must be a streaming model (
nemotron-3.5-asr-streaming-0.6borparakeet_realtime_eou_120m-v1); offline-only GGUFs failstream_begin. - Models are gitignored — you supply them per steps 3–4.
- App code: MIT (this repo,
LICENSE). - parakeet.cpp: MIT (submodule).
- Model weights follow each model's own license (e.g. NVIDIA nemotron is
OpenMDW-1.1; Moonshine per its repo). Apple
SpeechTranscriberis a system framework.
