Voz

Platforms

iOS, macOS, tvOS, visionOS, Android, Linux, Windows, Browser, Node

Languages

25

Weights

v0.4.0

Install

.package(url: "https://github.com/Desert-Ant-Labs/desert-ant-core.git", from: "4.0.0")

implementation("ai.desertant:voz:4.0.0")

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

Usage

Create one Voz instance and reuse it. The SDK downloads the model the first time you use it and keeps it on the device.

import Voz

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

result.text                     // the transcript
result.words.first?.start       // in seconds, 80ms resolution
result.realtimeFactor           // how many times faster than realtime

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

val voz = Voz(context)                       // downloads on first use
val t = voz.transcribe(samples, 16_000.0)

t.text                    // the transcript
t.words.first().start     // 80 ms resolution
voz.close()

let result = try await voz.transcribe(url, vocabulary: ["SoFi", "Timex", "Patek Philippe"])

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

if (!voz.isDownloaded()) voz.download()

let detection = try await Ear().identify(contentsOf: url)
guard detection.isReliable, Voz.supportedLanguages.contains(detection.language ?? "") else {
    return try await yourFallbackRecognizer(url)   // unreliable detection, or a language Voz doesn't support
}
let result = try await Voz().transcribe(url)

import { Voz } from "@desert-ant-labs/voz";

const voz = await Voz.load();
const result = await voz.transcribe(file);   // File, Blob, ArrayBuffer, or samples

result.text;
result.words[0];        // { text: "chapter", start: 0.08, end: 0.24 }
result.realtimeFactor;

Performance

We timed 10 minutes of audio on the Neural Engine and kept the fastest of three runs after the model loaded:

Short clips run slower than long files. On an M3 Ultra, a 5s clip takes 38ms (130x) and a 10s clip 48ms (210x).

On Apple platforms, Voz runs entirely on the Neural Engine, so your app keeps the GPU and CPU. On Windows, Voz runs on the GPU.

Accuracy

Word error rate

6.43% across six Open ASR Leaderboard datasets. Whisper large-v3-turbo scores 7.00%.

On long audio

2.48% on 5.4 hours of whole audiobook chapters.

Word timestamps

Starts are off by 80ms and ends by 77ms on average, compared with a forced aligner.

Size

337MB. Whisper large-v3-turbo is 1.6GB.

Voz scores better than Whisper large-v3-turbo at a fifth of the size. Voz scores four points better on meetings, about the same on podcasts and earnings calls, and worse on prepared financial speech.

Clean audiobook recordings score around 2%. Meetings, earnings calls and podcasts score 8-11%. Most real recordings sound more like those. On a podcast, plan to check one word in ten.

The model card has the accuracy for each language on long audio and the full leaderboard results.

Limits

• On Android and Linux, Voz runs through LiteRT, not through the generic inference path our other models share. The decode loop needs preallocated buffers and lane batching, which that path can't express.
• The speed figures above are from the Neural Engine. They don't transfer to LiteRT on a phone CPU. Measure on your target device before you promise real-time factors.
• In the browser, Voz needs its own 409MB download, because a GPU needs the weights in a different layout than the Neural Engine. Voz uses 1.2GB of memory in the browser, mostly for ONNX Runtime's compiled session. We test on Chromium 135+ and Safari 26+.
• In Node, Voz runs on the CPU. onnxruntime-node doesn't use the GPU by default, so a server transcribes slower than a browser on the same machine.
• Voz doesn't detect the language. Audio in a language Voz doesn't support comes back as fluent text that's wrong. Voz doesn't raise an error. Use Ear to check first.
• Accuracy depends on the language. Italian scores 5.05% and Greek 36.60% on the FLEURS test set. Check the model card before you promise a language to your users.
• Word ends are less precise than word starts. Voz marks where each word starts. The SDK estimates where each word ends from the audio. Timestamps land on 80ms frames, so no timestamp is more precise than 80ms.
• The model is 337MB. Download the model during onboarding, not on first use.