CoreAIKit

Transcribe — on-device speech-to-text, any ASR model in the catalog

The speech-to-text model runner: one app for every asr catalog entry — Whisper large-v3-turbo (100 languages), Qwen3-ASR 1.7B, Parakeet-TDT 0.6B v3, Fun-ASR-Nano 2512. Pick the model in the picker (it loads once and stays loaded), record a clip, choose a file, or use the bundled demo; transcription runs fully on device (first use downloads the model, then it loads from the local cache). Under the transcript, the app’s own measurement of the run: the clip’s length and how long the transcription call took, the model load not included.

let transcriber = try await KitTranscriber(catalog: "whisper-large-v3-turbo")
let samples = try AudioFile.pcm16kMono(url)  // any wav/m4a/mp3 → 16 kHz mono Float
let result = try await transcriber.transcribe(samples: samples)
// result.text, result.language

Run it

GUI (macOS / iOS):

open Transcribe.xcodeproj   # → Run, pick a model in the picker

CLI (macOS, headless — agents verify with this):

swift run transcribe-cli --audio sample.wav --model whisper-large-v3-turbo
swift run transcribe-cli --list-models

The transcript goes to stdout; download progress and the detected language go to stderr.

Hotwords — names and terms the model should spell right — go to Fun-ASR-Nano, comma-separated (in the app, the Hotwords field that appears when it is picked):

swift run transcribe-cli --audio memo.wav --model fun-asr-nano-2512 --hotwords "Tavenmoor, Anwen Thorsby"

Hands-off, for a recording or a smoke run: Transcribe.app/Contents/MacOS/Transcribe -autoplay 1 -model fun-asr-nano-2512 -clip <wav> -hotwords "…" -play 1 -log 1 loads the model, plays the clip, then presses Transcribe with the Hotwords field empty and again with it filled (Sources/Autoplay.swift has every flag).

The take-home: `Sources/QuickStart.swift`

No UI imports — a model you keep, and a one-call function over it:

enum SpeechModel { init(catalog id: String, …) async throws; func transcribe(samples:hotwords:…) }
func transcribe(audio url: URL, model id: String, hotwords: [String] = [], …) async throws -> Transcription

Both shells go through this file: the GUI view model holds a SpeechModel (loaded when the model is picked, kept between clips), the CLI calls transcribe(audio:model:), which loads the model per call. SpeechModel is KitTranscriber for every id except Fun-ASR-Nano, held as KitFunASRModel because only it takes hotwords. Want transcription in your own app? Copy that file — the glue (file → 16 kHz mono PCM, mic capture, id → engine dispatch) is all kit API (AudioFile.pcm16kMono, MicRecorder, KitTranscriber, KitFunASRModel), so it runs as pasted.

Integration checklist

SPM: github.com/john-rocky/coreai-kit → product CoreAIKit
Info.plist: NSMicrophoneUsageDescription (only if you record; the permission prompt and record button are app chrome — MicRecorder does the capture)
Entitlements (iOS): com.apple.developer.kernel.increased-memory-limit for Whisper (1.6 GB fp16 graph; 3.2 GB AOT bundle on iPhone)
First run downloads the model → cached in Application Support (progress callback)
Measure in Release — Debug is ~3× slower on per-token host work

Models

Catalog id	Model	Platforms	Notes
`whisper-large-v3-turbo`	Whisper large-v3-turbo	macOS + iOS	100 languages, auto-detect, ≤30 s window
`qwen3-asr-1.7b`	Qwen3-ASR 1.7B	macOS	LLM-engine ASR, 52 languages
`parakeet-tdt-0.6b-v3`	Parakeet-TDT 0.6B v3	macOS	TDT transducer, fastest of the three. Apple ships this model itself since `coreai-models` #136 (08-07), with live streaming in #184 — on iOS too. Reach for this entry when you want it inside the kit’s one-call ASR surface, not because it is otherwise unavailable
`fun-asr-nano-2512`	Fun-ASR-Nano 2512	macOS + iOS	zh / en / ja, hotword list in the prompt

This site is open source. Improve this page.