Key Takeaways
- Kokoro-82M: Apache-2.0, 82M parameters, runs via Apple's MLX framework through the community mlx-audio project β the real Apple Silicon acceleration path.
- Piper: GPL-3.0-or-later, CPU-only by design via ONNX Runtime β same performance profile on Intel and Apple Silicon Macs.
- XTTS v2: non-commercial CPML license, voice cloning from ~6 seconds of audio, but Metal (MPS) support is a tracked, unresolved GitHub issue β runs CPU-only on Mac.
- Bark: MIT license, experimental Apple Silicon MPS support behind an environment variable, no commits to its public repository since April 2024.
- Intel Macs cannot use the MLX-accelerated Kokoro path at all β MLX requires Apple Silicon.
π In One Sentence
For local text-to-speech on Apple Silicon, Kokoro-82M is the only one of the four major local engines with a real Metal GPU path (via the community mlx-audio project on Apple's MLX framework); Piper is the simplest CPU-only option that behaves identically on Intel Macs; XTTS v2 offers voice cloning but its Apple Silicon MPS support is a documented broken GitHub issue; and Bark's MPS support is explicitly experimental.
π¬ In Plain Terms
Not every free text-to-speech program you can run on a Mac actually uses the Mac's graphics chip β some just run on the regular processor, which works fine but is slower than it could be. Kokoro is the one built to take advantage of Apple's own chip design through a project called MLX; the others either skip the GPU by design (Piper) or try to use it but run into unresolved problems (XTTS v2, Bark).
πNote: This guide compares the four local TTS engines PromptQuorum has independently reviewed (Piper, Kokoro, XTTS v2, Bark). Each has its own dedicated review linked throughout for install commands, licensing detail, and full limitations.
Which Local TTS Engine Should You Actually Install?
The right engine depends on whether you need GPU acceleration, voice cloning, or the widest possible Mac compatibility β no single engine wins on all three. Kokoro is the only engine here with a genuine Apple Silicon GPU path; Piper is the most portable; XTTS v2 is the only one that clones voices.
- π Best overall for Apple Silicon: Kokoro-82M via mlx-audio β the only engine covered here with a real Metal GPU path, small enough (82M parameters) to run comfortably on any Apple Silicon Mac.
- Best for simplicity and Intel compatibility: Piper β CPU-only by design, so install and performance are identical whether you are on an M-series chip or an older Intel Mac.
- Best if you need voice cloning: XTTS v2 β clones a voice from ~6 seconds of reference audio, but runs CPU-only on any Mac and is licensed for non-commercial use only.
- Best for expressive, non-speech audio: Bark β laughter, sighs, and simple music from text, with experimental (partial) Apple Silicon GPU support, though its repository has been dormant since April 2024.
Who Should Use Which Engine?
Match the engine to your Mac and your actual requirement, not to whichever one has the most GitHub stars. GPU acceleration only matters if you are generating enough audio, often enough, for CPU-only synthesis to feel slow.
- π§ Apple Silicon Mac, want the fastest local option: Kokoro via mlx-audio β the only engine here written to use Metal through MLX.
- π§ Any Mac, including an older Intel model: Piper β CPU-only by design, so there is nothing Apple-Silicon-specific to configure or troubleshoot.
- π§ Need to clone a specific voice from a short recording: XTTS v2 β accept that it will run on CPU only on a Mac, and that its license is non-commercial.
- π§ Want laughter, sighs, or ambient sound, not just speech: Bark β but budget extra install time for its experimental Apple Silicon path, and confirm current maintenance status first.
- β Skip Bark if you need active maintenance or guaranteed performance β its public repository has shown no commits since April 2024, independent of the Apple Silicon question.
- β Skip XTTS v2 if you are building a commercial product β its Coqui Public Model License (CPML) is non-commercial, and Coqui AI, the company that released it, shut down its paid services in December 2023.
How Do Piper, Kokoro, XTTS v2, and Bark Compare on Apple Silicon Fitness?
Kokoro is the only engine in this comparison with a genuine, purpose-built Apple Silicon GPU path; the rest are CPU-only by design or by unresolved bug. The table below scores each engine on the four factors that actually decide Mac fitness: whether it uses Apple Silicon acceleration, how much memory it needs, how much friction the macOS install has, and voice quality.
Engine | Apple Silicon accel | RAM footprint | macOS install | Voice quality |
|---|---|---|---|---|
| Kokoro-82M | Yes β Metal via MLX (mlx-audio) | Low (82M params, quantized variants exist) | pip install + mlx-audio, Apple Silicon only | Natural, close to larger cloud models |
| Piper | None by design β CPU-only ONNX Runtime | Very low (~50β100MB per voice) | pip install piper-tts, no GPU setup | Clear, some robotic prosody |
| XTTS v2 | None β MPS hangs (GH issue #3649) | Moderateβhigh (full cloning model) | pip install coqui-tts, CPU-only on Mac | High, clones a specific voice |
| Bark | Experimental β SUNO_ENABLE_MPS=True | High (use small-models flag to reduce) | pip install from GitHub, no PyPI package | Expressive, non-deterministic |
Voice-quality ratings are qualitative descriptions based on each engine's documented architecture and PromptQuorum's dedicated reviews (linked per row), not a blind listening test PromptQuorum conducted β no numeric MOS (Mean Opinion Score) or benchmark figures are claimed here.
Which Engines Actually Use the GPU on a Mac?
Only Kokoro has a real, purpose-built path to Apple Silicon's Metal GPU; Piper skips the GPU entirely by design, and XTTS v2 and Bark both have unresolved or partial GPU support on Mac. This is the single biggest factor separating these engines on Apple Silicon, and it is easy to assume "runs on a Mac" means "uses the Mac's GPU" β it usually does not.
- Kokoro-82M runs through Apple's own MLX framework via the community project mlx-audio, which requires Apple Silicon and Python 3.10β3.12. MLX is Apple's open-source machine learning framework, built from the ground up for Metal on Apple Silicon's unified memory architecture β the same framework the site's Apple Silicon local LLM guide and MLX vs. Ollama vs. llama.cpp comparison cover for language models. The official Kokoro-82M weights, released by hexgrad on Hugging Face, are natively a PyTorch model; the MLX path is a community port, not an official Apple or hexgrad release, and mlx-audio also offers quantized (bf16, 8-bit, 4-bit) variants for lower memory use.
- Piper never touches the GPU on any platform, Apple Silicon included β this is by design, not a limitation. Piper converts text to phonemes with espeak-ng, then synthesizes audio with a model exported to ONNX Runtime for fast CPU inference. That design choice is exactly why Piper runs in real time even on a Raspberry Pi β see PromptQuorum's Piper TTS review for the full architecture and install steps.
- XTTS v2's Apple Silicon Metal (MPS) support does not currently work. A tracked issue on the coqui-ai/TTS GitHub repository, titled "Unable to use xtts_v2 with mps device on Apple Silicon," documents that attempting to run XTTS v2 on the MPS device hangs rather than completing. Coqui's own project does not list Apple Silicon GPU support as officially supported. In practice, this means XTTS v2 runs on CPU only on a Mac, through the same Coqui TTS toolkit (MPL-2.0 licensed) that runs it on other platforms.
- Bark has experimental Apple Silicon MPS support, gated behind an environment variable. Setting
SUNO_ENABLE_MPS=Trueenables Metal acceleration, per discussion on the suno-ai/bark GitHub repository, but some PyTorch operators Bark depends on were not implemented for MPS at the time, causing partial fallback to CPU for those steps. Bark also supports aSUNO_USE_SMALL_MODELS=Trueflag specifically to reduce memory pressure on Macs with less unified memory.
How Do You Install Kokoro with MLX Acceleration on a Mac?
This walkthrough installs the community mlx-audio project to run Kokoro-82M through Apple's MLX framework, using the setup documented in the mlx-audio GitHub repository.
- 1Confirm you are on Apple Silicon with a supported Python version.
Why it matters: mlx-audio requires an Apple Silicon Mac (M-series chip) and Python 3.10β3.12; MLX does not run on Intel Macs at all, so this path is Apple-Silicon-only. - 2Install mlx-audio.
Why it matters: Run `pip install mlx-audio` in a Python virtual environment. This pulls in MLX itself (version 0.31 or later) alongside the audio pipeline. - 3Run a first synthesis from the command line.
Why it matters: The package ships a CLI entry point that downloads the Kokoro-82M weights on first use and synthesizes a WAV file from a text string β check the project's current README for the exact command, since CLI flags can change between releases. - 4Pick a voice and, optionally, a quantized model variant.
Why it matters: Kokoro-82M ships with 54 voice presets across several languages. mlx-audio also offers bf16, 8-bit, and 4-bit quantized variants β lower precision trades a small amount of quality for a smaller memory footprint, useful on a Mac with less unified memory. - 5Integrate into your own Python application.
Why it matters: For anything beyond one-off CLI synthesis, call mlx-audio's Python API directly rather than shelling out to the CLI repeatedly, avoiding the model-reload cost on every call.
What Changes on an Intel Mac?
On an Intel Mac, the MLX-accelerated Kokoro path is unavailable entirely β MLX requires Apple Silicon and will not run at all on Intel hardware. Every other engine covered here still works on Intel, because none of them depend on Apple's Neural Engine or Apple Silicon-specific GPU acceleration to function; they simply run on CPU.
- Piper is unaffected by the Intel/Apple Silicon distinction. It is CPU-only by design on every platform, so an Intel Mac performs comparably to an Apple Silicon Mac for Piper specifically, hardware generation aside.
- Kokoro still runs on an Intel Mac through its official PyTorch weights, just without the MLX acceleration path. You lose the Apple Silicon-specific Metal route via mlx-audio, but the model itself (82M parameters) is small enough to run acceptably on CPU.
- XTTS v2 and Bark run identically on Intel and Apple Silicon Macs, since both currently run CPU-only on any Mac anyway β XTTS v2 because MPS support is broken, and Bark because MPS support is experimental and partial. Neither loses meaningful capability moving from Apple Silicon to Intel, since neither has a mature accelerated path on Apple Silicon to begin with.
When Should You Not Use Any of These Engines?
None of the four engines in this comparison are the right choice for every Mac TTS use case β each has situations where a different tool, or a cloud API, fits better.
- β Need a guaranteed commercial license with cloned voices. XTTS v2's CPML license is non-commercial, with no confirmed active path to a commercial license since Coqui AI shut down its paid services in December 2023 β see PromptQuorum's ElevenLabs comparison for a managed commercial cloud alternative.
- β Need guaranteed active maintenance. Bark's public GitHub repository has shown no commits since April 5, 2024; if ongoing fixes and updates matter for your project, Piper (actively maintained by the Open Home Foundation) or Kokoro (an actively used community ecosystem around mlx-audio) are safer bets.
- β Need production-grade Apple Silicon GPU support today, with no community-project dependency. Kokoro's MLX path runs through a community project, not an official Apple or hexgrad release β treat it as good, but not vendor-guaranteed.
- β Need real-time interactive voice on very limited Apple Silicon memory (8GB base configurations) while also running a large local LLM at the same time. Stacking a large XTTS v2 or Bark process alongside an LLM on a memory-constrained Mac can be tight; Piper's and Kokoro's small footprints leave more headroom.
Frequently Asked Questions
What is the best local TTS engine for Apple Silicon Macs?
Kokoro-82M, run through the community mlx-audio project, is the best choice if you specifically want Apple Silicon GPU (Metal) acceleration via Apple's own MLX framework. If you want the simplest install that works the same on any Mac, Piper is the better pick, since it is CPU-only by design on every platform.
Does Piper use the GPU on a Mac?
No. Piper is CPU-only by design on every platform, including Apple Silicon and Intel Macs. It converts text to phonemes with espeak-ng and synthesizes audio through ONNX Runtime, which is why it runs in real time even on a Raspberry Pi without any GPU at all.
Can Kokoro-82M run on Apple Silicon with GPU acceleration?
Yes, through the community project mlx-audio, which runs Kokoro-82M via Apple's own MLX framework β built specifically for Metal on Apple Silicon's unified memory architecture. The official Kokoro-82M weights from hexgrad are a PyTorch model; the MLX path is a community port, not an official release, and requires an Apple Silicon Mac (Intel Macs cannot use it) and Python 3.10β3.12.
Does XTTS v2 work on Apple Silicon?
It runs, but CPU-only. XTTS v2's Metal (MPS) device support is a documented, unresolved issue tracked on the coqui-ai/TTS GitHub repository (issue #3649), where attempting MPS device use hangs rather than completing. Coqui's project does not officially support Apple Silicon GPU acceleration for XTTS v2, so expect CPU-only performance on any Mac.
Is Bark accelerated on Apple Silicon?
Partially, and experimentally. Setting the SUNO_ENABLE_MPS=True environment variable enables Metal GPU acceleration for Bark, but some PyTorch operators it depends on have not been implemented for MPS, so some processing steps still fall back to CPU. Bark's own maintainers describe this support as experimental, not production-ready.
Can I use any of these engines on an Intel Mac?
Piper, XTTS v2, and Bark all run on Intel Macs, since none of them require Apple Silicon-specific acceleration to function β they run on CPU either by design (Piper) or because their GPU acceleration paths are unresolved or partial anyway (XTTS v2, Bark). Kokoro's MLX-accelerated path specifically requires Apple Silicon and will not run on an Intel Mac at all, though Kokoro's official PyTorch weights still run on Intel without MLX acceleration.
Which of these engines can clone a specific voice?
Only XTTS v2, among the four covered here, clones a voice from a short reference audio clip (as little as 6 seconds, per its official model card). Piper, Kokoro, and Bark all use pre-trained or preset voices rather than cloning an arbitrary voice on the fly. See PromptQuorum's dedicated XTTS v2 review for full cloning details and licensing.
Do any of these local TTS engines require a paid license for Apple Silicon Mac use?
No engine covered here charges specifically for macOS or Apple Silicon use. Piper (GPL-3.0-or-later), Kokoro (Apache-2.0), and Bark (MIT) are all free and open-source software regardless of platform. XTTS v2 is free to use but under a non-commercial license (CPML) β that restriction applies equally on Apple Silicon, Intel, Windows, or Linux, and is unrelated to which Mac you use it on.
Verdict
On Apple Silicon specifically, Kokoro-82M stands apart because it is the only one of these four engines with a genuine, purpose-built path to the Mac's Metal GPU, through the community mlx-audio project built on Apple's own MLX framework β and it is small enough (82 million parameters, Apache-2.0 licensed) that this acceleration is worth setting up. Piper remains the right default when simplicity and cross-hardware consistency matter more than raw speed: it is CPU-only everywhere, so there is nothing Apple-Silicon-specific to configure, troubleshoot, or worry will regress on an Intel Mac. XTTS v2 is worth the CPU-only performance hit only if you specifically need voice cloning and can live with its non-commercial license; Bark is worth considering only for its distinctive non-speech sounds, with the caveat that both its Apple Silicon acceleration and its overall maintenance status are unsettled. If unsure, start with Piper for the simplest install, move to Kokoro via mlx-audio once you confirm you want the Metal acceleration, and reach for XTTS v2 only when voice cloning is a hard requirement.
Sources
- Kokoro-82M on Hugging Face β the official model card: parameters, license, and architecture.
- mlx-audio on GitHub β the community project that runs Kokoro-82M via Apple's MLX framework on Apple Silicon.
- coqui-ai/TTS GitHub issue #3649 β "Unable to use xtts_v2 with mps device on Apple Silicon," documenting the unresolved MPS hang.
- suno-ai/bark GitHub repository β issues and pull requests documenting experimental Apple Silicon MPS support via SUNO_ENABLE_MPS.
- Apple MLX Framework β Apple's official open-source machine learning framework with native Metal GPU acceleration for Apple Silicon.
- Piper TTS Review β PromptQuorum's dedicated review, including install commands and license history.
- XTTS v2 Review, Coqui TTS Review, and Bark TTS Review β PromptQuorum's dedicated reviews of the other engines covered here.