WhisperKit
On-device speech recognition for Apple Silicon: OpenAI Whisper running locally in Swift, no cloud round-trip.
About
WhisperKit brings OpenAI's Whisper speech-to-text to Apple devices as a native Swift framework, running entirely on-device via Core ML. No audio leaves the machine, latency is local, and it works offline, which makes it the natural choice for iOS and macOS apps that need transcription without a server bill or a privacy review.
It now lives inside Argmax's open-source Swift SDK alongside SpeakerKit (speaker diarization with Pyannote) and TTSKit (text-to-speech with Qwen-TTS): one package gives you a full on-device speech pipeline. The repo also ships a Swift CLI and a local server with a generated API spec, so you can prototype from the terminal before embedding it in an app. Pre-converted Core ML Whisper models are published on the Hugging Face Hub. At 6,400+ stars and MIT it is the reference on-device STT stack for the Apple ecosystem. The boundary is the platform: this is Apple Silicon first, and model quality follows whichever Whisper checkpoint you load.
What it can do
- On-device Whisper speech-to-text via Core ML, fully offline
- SpeakerKit: speaker diarization with Pyannote models
- TTSKit: text-to-speech with Qwen-TTS
- Pre-converted Core ML models published on the Hugging Face Hub
- Swift CLI and local HTTP server for prototyping and testing
- Memory-efficient loading for large audio files; MIT licensed
Setup
- Add the package via Swift Package Manager: github.com/argmaxinc/argmax-oss-swift (Xcode or Package.swift).
- Download a pre-converted Core ML Whisper model from huggingface.co/argmaxinc/whisperkit-coreml, or generate your own per the docs.
- Transcribe with the WhisperKit API (see the Quick Example in the README); add SpeakerKit for diarization or TTSKit for speech output.
- For prototyping, build the Swift CLI or run the local server target instead of embedding in an app.
Prerequisites
Apple Silicon Mac or iOS device; Xcode / Swift toolchain for app integration; a Core ML Whisper model from the Hub or self-converted.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: 6,404 GitHub stars (GitHub API), MIT, pushed Oct 11 2026; the README documents WhisperKit/SpeakerKit/TTSKit, the SPM install, the Swift CLI and local server; the Hugging Face Hub corroborates the published Core ML models. Strengths: the reference on-device STT stack for Apple, full speech pipeline in one SDK, genuinely offline. Docked for: (1) Apple Silicon first, limited elsewhere; (2) WhisperKit recently folded into the argmax-oss-swift monorepo, docs in transition; (3) quality follows the chosen Whisper checkpoint. License: MIT.
★ 6,404 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Adoption 6: starScore from 6,404 stars on argmaxinc/argmax-oss-swift (log scale vs 298k peer max)
- · Maintenance 9: pushed 2026-10-11 (recency 10); docked one point for monorepo transition
- · Openness 9: MIT (0.7*10) + HF Hub Core ML models verified (0.3*10); docked: docs in transition after WhisperKit folded into argmax-oss-swift
- · Manual 8: reference on-device STT for Apple corroborated by README and Hub models; docked: Apple Silicon first, quality follows chosen checkpoint
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.