Core AI model zoo

Whisper large-v3-turbo — Core AI

openai/whisper-large-v3-turbo (809M, MIT) — OpenAI’s multilingual speech-to-text: 100 languages, automatic language detection, ≤30 s window per decode. Runs fully on device on the stock Core AI runtime (no engine patch), token-for-token identical to the PyTorch greedy reference.

Bundle: 🤗 mlboydaisuke/whisper-large-v3-turbo-CoreAI-official — macOS (~1.6 GB fp16) + iOS (~3.2 GB, AOT-precompiled for iPhone). Catalog id: whisper-large-v3-turbo.

Whisper large-v3-turbo demo Whisper large-v3-turbo on iPhone 17 Pro — the zoo’s coreai-audio app, real speed.

Use it

⚡ One line — this model is the default behind the kit’s task op (import CoreAIOps; no session, no model plumbing, downloads on first use):

let text = try await CoreAI.transcribe(audioURL)

Twenty ops, one shape — Cookbook.

▶️ Run it (source) — the Transcribe runner (GUI + CLI, one app for every speech-to-text model in the catalog):

git clone https://github.com/john-rocky/coreai-kit
open coreai-kit/Examples/Transcribe/Transcribe.xcodeproj
# → Run, then pick "Whisper large-v3-turbo" in the model picker

# agents / headless (macOS):
cd coreai-kit/Examples/Transcribe
swift run transcribe-cli --model whisper-large-v3-turbo --audio sample.wav

💻 Build with it — complete; the glue is kit API, copy-paste runs:

import CoreAIKit

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 ("en", "ja", … auto-detected)

The take-home is Examples/Transcribe/Sources/QuickStart.swift — this exact code as one typed function, no UI; both the runner’s GUI and its CLI call it. Recording? MicRecorder (kit API) captures mic audio as 16 kHz mono [Float] — the record button and permission prompt are your app’s own chrome.

Integration checklist

Measured (M4 Max, GPU)

Greedy decode vs the HF PyTorch reference (generate, greedy):

Metric Value
Transcript token-for-token identical to PyTorch greedy
First step (compile + warmup) 0.68 s
Per token (steady state) 0.18 s

The fixed window caps one decode at 128 tokens — enough for a 30 s window; chunk longer audio into 30 s segments (the runner and KitWhisperModel do this for you).

How the bundle works (why it’s not the stock recipe verbatim)

Apple’s official models/whisper/export.py traces a single decode step (decoder_input_ids [1,1], no KV cache) — that graph can’t be driven autoregressively. This bundle is the same recipe traced at a fixed 128-token decoder window: pad the buffer, read the logits at the real last position (causal attention ignores the padding, so the read is exact), constant shape so MPSGraph compiles once (a dynamic-length export recompiles every step → ~15 s/token; this is 0.18 s/token). Full write-up: knowledge/whisper-asr-fixed-decode.md · recipe: conversion/export_whisper_fixed.py · hashes + export environment: the HF card.

Driving the raw graph yourself (no kit)? apps/CoreAITranscribe implements this exact decode loop + the log-mel frontend from scratch (macOS + iOS, file or mic) — the reference if you’re porting to another language or runtime.

This site is open source. Improve this page.