Transcribing a file

The shared instanceCopy Anchor Link

Scribe is an actor that holds the loaded models. Use one instance for the whole app, so the models load once:

let scribe = Scribe.shared

Scribe() creates a separate instance with its own models, which is rarely what you want.

Plain textCopy Anchor Link

For dictation-style use, when you only need the words:

let text = try await scribe.transcribeText(url, language: "pl")

words(_:language:) returns the same result as timed Word values.

With speakersCopy Anchor Link

let transcript = try await scribe.transcribe(url, options: .init(language: "en"))

for segment in transcript.segments {
    let name = transcript.speaker(for: segment.speakerID)?.name ?? "Unknown"
    print("[\(TranscriptRenderer.clock(segment.start))] \(name): \(segment.text)")
}

Diarization is on by default. It costs roughly one extra pass over the audio.

The input is any file AVAudioFile can read, for example WAV, M4A, MP3 or CAF.

OptionsCopy Anchor Link

TranscriptionOptions controls both calls:

let options = TranscriptionOptions(language: "de", speakerCount: 3,
                                   remoteSpeakerTemplate: "Sprecher %d")

LanguagesCopy Anchor Link

Parakeet v3 detects the language on its own. The hint mainly stops words from being written in the wrong script, for example Polish words in Cyrillic. An unknown code throws ScribeError.unsupportedLanguage.

Scribe.supportedLanguages lists the codes the hint accepts: bg, cs, da, de, el, en, es, et, fi, fr, hr, hu, it, lt, lv, mt, nl, pl, pt, ro, ru, sk, sl, sv and uk, plus be, bs and sr.

SpeakersCopy Anchor Link

Speakers get stable IDs (s1, s2, …) numbered in the order they first speak. The diarizer's own labels are arbitrary, so ScribeKit renumbers them, and speakers that end up with no words are dropped.

Rename them once you know who is who:

var transcript = try await scribe.transcribe(url)
transcript.rename(speaker: "s1", to: "Anna")

Each word goes to the speaker whose turn overlaps it most, or to the nearest turn when none does. A new segment starts when the speaker changes, after a pause of 2 seconds, or at the first sentence end after 45 seconds.

ProgressCopy Anchor Link

Pass a closure to follow the stages:

let transcript = try await scribe.transcribe(url) { stage in
    switch stage {
    case .loadingModels: print("Loading models")
    case .transcribing(let track): print("Transcribing \(track)")
    case .identifyingSpeakers: print("Finding speakers")
    case .finishing: print("Finishing")
    }
}

The closure is called from the Scribe actor. Hop to the main actor before touching UI.

Diarization onlyCopy Anchor Link

diarize(_:speakerCount:) returns speaker turns without transcribing, for example to combine with another recognizer:

let turns = try await scribe.diarize(url)
let segments = SegmentBuilder.segments(words: myWords, turns: turns)