What does Supertonic 3 make possible on a Mac without a discrete GPU?

By Wei Jie Chee · 22 May 2026, 00:00 Z

Download printable cheat-sheet (CC-BY 4.0)
Constrained implementer + Named-model investigatorFit hardware, latency, rights, and workflow contracts together.

The tested ONNX path generated speech on Apple Silicon CPU, creating a credible edge lane with different quality and cloning expectations.

Do this first

Check whether preset voices and CPU latency meet the job before treating voice cloning or adaptation as requirements.

Keep this boundary in view

One macOS smoke test does not establish every Mac, browser, language, long-form workload, or listener preference.

Inspect the evidence and what would change the answer

First-party Apple Silicon smoke test

What supports it

The page records setup, model cache, memory, output, CPU behavior, and comparison boundaries against larger models.

Change our mind if

Repeated target-device tests fail the latency or quality threshold, or a smaller route performs better under the same constraints.

Claim TTS-SUPERTONIC-MAC-2026-08 / v1, checked 7 Aug 2026

Supertonic 3 is interesting for a different reason than most new TTS releases.

It is not trying to win the biggest-model contest. It is a compact, ONNX Runtime based TTS system designed to run locally across desktop, browser, mobile, and edge targets. That makes it relevant for product teams that care about offline use, privacy, cold-start time, and simple deployment more than maximum voice-clone fidelity.

We ran a bounded macOS smoke test to answer the practical question first: can Supertonic 3 run locally on Apple Silicon, how much memory does it use, and what kind of output does the default Python SDK produce?

60-second takeaway

  • It runs locally on macOS. The Python SDK installed cleanly in a Python 3.12 uv environment and generated a valid WAV on Apple Silicon.
  • The default path is CPU-first. ONNX Runtime exposed CoreML on this host, but Supertonic's package default used CPUExecutionProvider.
  • Memory use was modest for local TTS. The cached short synthesis peaked at about 547 MB resident memory. Model initialization alone peaked at about 528 MB.
  • No discrete VRAM was used. On Apple Silicon, GPU and CPU share unified memory, but this test did not use a GPU execution provider.
  • This is an on-device deployment candidate, not a replacement for fine-tuned voice-cloning models. Put it next to Kokoro in the edge lane, not next to IndexTTS2 or VoxCPM full-SFT.

What Supertonic 3 is

The upstream project describes Supertonic as an on-device multilingual TTS system powered by ONNX Runtime. The public repository lists Python, Node.js, browser, Java, C++, C#, Go, Swift, iOS, Rust, and Flutter examples, plus a Python quick start that auto-downloads model assets on first run.

Supertonic 3, released on April 29, 2026, adds:

  • 31-language support
  • public ONNX assets compatible with the v2 code path
  • fewer repeat and skip failures, according to upstream release notes
  • 44.1 kHz 16-bit WAV output
  • inline expression tags
  • a compact open-weight model positioned around 99M parameters

The model weights are released separately from the sample code. Treat the code license and model license as separate production checks.

Our macOS smoke test

We tested the Python SDK path because it is the shortest path from zero to local audio output.

Host and environment

Item