voz

Transcribe 10 minutes in 2 seconds.

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

• SDKs, install and examples: https://desertant.com/docs/voz/

Voz transcribes speech on the device, in 25 languages, with word-level timestamps. You can run Voz on iPhone, iPad, Mac, Windows, in the browser and in Node, so your audio never leaves the device. An iPhone 17 Pro transcribes 10 minutes of audio in 1.8s.

Try it

Voz is part of our SDK, Desert-Ant-Labs/desert-ant-core.

• Apple platforms and Windows: the Swift SDK, through Swift Package Manager.
• Browser and Node: the JavaScript SDK, npm i @desert-ant-labs/voz, plus onnxruntime-web in the browser or onnxruntime-node in Node.

The SDK downloads the model on first use and caches the files, so your app stays small.

let voz = try await Voz()
let result = try await voz.transcribe(url)
result.text                       // "We made the bet to build for iOS..."
result.words.first?.start         // 0.32
result.words.first?.end           // 0.58

Install

Swift (requirements)

.package(url: "https://github.com/Desert-Ant-Labs/desert-ant-core.git", from: "4.2.0")

Then add the Voz product to your target.

Kotlin (requirements)

implementation("ai.desertant:voz:4.2.0")

JavaScript (requirements)

npm i @desert-ant-labs/voz onnxruntime-web

Files

We ship the Core ML models compiled. If you turn them back into an .mlpackage, Core ML recompiles them on every launch and a load takes 28s instead of 0.13s.

How it works

Voz is NVIDIA's Parakeet TDT 0.6B v3, converted and compressed to run on phones, laptops and in the browser. Voz cuts a recording into 15s windows at pauses, transcribes up to 16 windows at once, and joins them where neighboring windows agree on the words. Because Voz works through a recording window by window, a five-hour file needs no more memory than a five-minute one. On Apple devices every step runs on the Neural Engine.

You can pass audio at any sample rate, mono or stereo. The SDK handles the conversion.

Speed

Time to transcribe 10 minutes of audio with Swift SDK 3.5.0, the fastest of three runs after the model loaded:

Short clips run at 50-62x, because Voz always processes a full 15s window.

The first load after install takes 20s while Core ML prepares the model for the Neural Engine. After that, a load takes 0.2s on an M3 Ultra.

Accuracy

English

We scored Voz on the Open ASR Leaderboard datasets with the leaderboard's own text normalizer. Whisper's numbers come from the leaderboard.

Voz comes close to Whisper large-v3-turbo at less than a third of the size. Voz does better on meetings and worse on read and prepared speech.

Clean audiobook recordings score 2-4%. Meetings, earnings calls and podcasts score 10-13%, and most real recordings sound more like those. On a podcast, plan to check one word in ten.

We left out VoxPopuli and TEDLIUM. The leaderboard's scripts point at a smaller VoxPopuli set than the one behind its published scores, and the TEDLIUM config has no data.

Word timestamps

We compared Voz's word times with torchaudio's MMS_FA forced aligner.

Word ends are harder to get right. The model reports how long each token lasts, and that tends to run into the pause after a word, so the SDK trims the end back using the audio. The aligner isn't exact either, and Voz works in 80ms frames, so no timestamp is more precise than 80ms.

Accuracy by language

We tested each language on 10 minutes of audio, made by joining FLEURS test clips into one long file. A single FLEURS clip is shorter than one 15s window, so a single clip wouldn't test how Voz joins windows, which every real recording needs.

|