Voz

Every word, with the time it was said.

On-device speech recognition: transcripts with word-level timestamps, 25 languages.

Install

Swift (requirements)

.package(url: "https://github.com/Desert-Ant-Labs/desert-ant-core.git", from: "3.1.0")

Then add the Voz product to your target.

Usage

Voz turns speech into text, with a start and an end on every word. Create one and reuse it; the model downloads on first use and is cached.

import Voz

let voz = try await Voz()
let result = try await voz.transcribe(url)

result.text                     // the transcript
result.words.first?.start       // 80 ms resolution
result.realtimeFactor           // seconds of audio per second of wall clock

Samples work too, mono at voz.sampleRate:

let result = try await voz.transcribe(samples: samples)

Downloading ahead of time

The first load after a download pays a one-time Neural Engine specialization of roughly 20 seconds; every load after it takes about 0.2 s. Doing both during onboarding keeps that cost off the first transcription.

if !Voz.isDownloaded() {
    try await Voz.download { progress in
        show(progress.fraction)
    }
}

Picking a language first

Voz covers 25 languages and does not detect which one it is hearing. Pair it with Ear when the input could be anything:

let detection = try await Ear().identify(contentsOf: url)
guard detection.isReliable, Voz.supportedLanguages.contains(detection.language ?? "") else {
    return try await yourFallbackRecognizer(url)   // Voz does not cover it
}
let result = try await Voz().transcribe(url)

The fallback is yours to choose: Voz ships the recognizer, not a router.

Accuracy

Close to a model three and a half times its size, two points better on meetings, behind on prepared and read speech.

Expect the conversational figures, not the LibriSpeech one. Read speech in a clean recording scores around 2%; meetings, earnings calls and podcast audio score 10-13%, and most real material is nearer the second group. Roughly one word in ten wanting a look is the honest expectation for a podcast.

Per-language figures on long audio, and the full leaderboard breakdown, are in the model card.

Limits

• Apple platforms only. The runtime drives Core ML directly, because the things that make it fast (preallocated buffers, outputBackings, a lane-batched decode loop) are not expressible through the generic inference shape the other models share. There is no Android, Linux or web build.
• 25 languages, and it does not know which one it is hearing. Feeding it a language it does not cover produces confident nonsense rather than an error. See Ear.
• Accuracy varies widely by language. Italian is 3.31% and Greek 39.46% on the same ten-minute-per-language protocol. Check the model card before promising a language.
• Word ends are the harder half. The recognizer reports how far to skip after each token rather than where a word stops, so ends are trimmed back using the audio. 80 ms is the frame resolution and the floor for any timestamp here.
• 467 MB is a real download. Fetch it during onboarding, not on first use.