Introduction

ScribeKit is a Swift package that turns an audio file into a transcript with speakers. It wraps FluidAudio's Core ML port of NVIDIA Parakeet TDT v3 and its offline speaker diarization in a small API, and renders the result as Markdown, plain text, SRT, WebVTT or JSON.

Everything runs on the Mac or iPhone that calls it. There is no server, API key or Python runtime.

import ScribeKit

let transcript = try await Scribe.shared.transcribe(url, options: .init(language: "en"))
print(TranscriptRenderer.markdown(transcript))

What it doesCopy Anchor Link

• Speech recognition with Parakeet TDT 0.6B v3: 25 European languages, detected automatically, with word timings.
• Speaker diarization: who spoke when, with speakers numbered in the order they first talk.
• Conversation mode for calls recorded as two tracks, the microphone and the system audio. The microphone is always the local speaker, and only the other side is diarized.
• Echo removal: words the microphone picked up from the speakers are dropped, so the other side is not transcribed twice.
• Readable segments, split on speaker changes and pauses.
• Renderers for Markdown with YAML front matter, plain text, SRT, WebVTT and JSON.

The segment builder and the renderers are plain Swift with no models, so you can also feed them words and speaker turns from another recognizer.

RequirementsCopy Anchor Link

Where nextCopy Anchor Link