F5-TTS on Mac: Setup and Alternatives Compared

Direct answer: F5-TTS is an open-source, zero-shot text-to-speech model that you can run on an Apple Silicon Mac in two main ways: the official PyTorch package installed with pip, or a community MLX implementation built for Apple silicon. It is genuinely useful for research and voice-cloning experiments, and its code is MIT licensed, but its pretrained weights are released under a non-commercial license, which rules it out of most paid client work. If you want finished, sellable narration without a Python environment, a packaged local Mac app is the more practical choice.

Disclosure: we build Murmur, a one-time Mac voice studio, so we have a stake in the no-code route. F5-TTS itself is not one of the models Murmur ships; it is a separate open-source project. Everything below about F5-TTS comes from its own repository and model card, and we did not measure its audio quality. Confirm the current license and behavior yourself before you rely on the output.

What F5-TTS is, in plain terms

F5-TTS is described by its authors as a flow-matching model for text to speech, and the repository calls itself a fairytaler that fakes fluent and faithful speech. In practice it is non-autoregressive: instead of generating audio token by token, it predicts a mel spectrogram with a diffusion transformer and then vocodes it, which lets it clone a voice zero-shot from a short reference clip plus that clip's transcript. It is an evolution of E2 TTS, adding ConvNeXt V2 blocks for text alignment. The official GitHub repository listed about 15.4k stars when we checked, and the Hugging Face model card showed roughly 822,000 downloads in the last month, which tells you how many people are experimenting with it.

Two details shape the Mac experience before you install anything. First, the model was trained on Emilia, which the authors describe as an in-the-wild dataset, and that is why the weights carry a non-commercial license even though the code is MIT. Second, the weights live separately from the code: the model card tells you to download F5TTS_Base, F5TTS_v1_Base, or E2TTS_Base and place them under a checkpoints folder, so installing F5-TTS is really two steps, the package and the checkpoint.

Three ways to run F5-TTS on a Mac

There is no single Mac installer, so the practical question is which route matches your comfort with a terminal.

Route 1: the official PyTorch package

The repository documents a conda environment with Python 3.10 or newer, an FFmpeg install, and, for Apple Silicon, a plain pip install torch torchaudio. You then install the package with pip install f5-tts. From there you can launch the Gradio web interface with f5-tts_infer-gradio, or run the command line directly with f5-tts_infer-cli, passing a model name, a reference wav path, the reference transcript, and the text you want spoken. The README notes that if you leave the reference text empty, an automatic speech recognition model transcribes it for you at extra memory cost. This is the most faithful route: you are running the authors' own code, and it also includes training and fine-tuning tools.

Route 2: an MLX port for Apple silicon

If you would rather stay in Apple's own machine-learning stack, the MLX implementation is the one the F5-TTS repository acknowledges. It installs with pip install f5-tts-mlx and runs as python -m f5_tts_mlx.generate. For voice matching it asks for a mono, 24 kHz wav of roughly 5 to 10 seconds plus its transcript, and it supports 4-bit and 8-bit quantized weights with a --q option for memory-limited machines. The author reported a sample generated in about 4 seconds on an M3 Max, which is their measurement and not ours. If you want the background on why MLX matters on a Mac at all, our what is MLX TTS explainer covers it.

Route 3: a packaged app, where F5-TTS is not

The no-code route on a Mac is a packaged voice app, and here is the honest part: F5-TTS is not one of the models Murmur ships. Murmur bundles Kokoro and can install others such as Chatterbox Turbo and Chatterbox Multilingual, Qwen3-TTS Base, Voice Design, SparkTTS, Breeze TTS 2, Fish Audio S2 Pro, and OmniVoice, with a reference-cloning workflow, voice design, multi-speaker projects, and export. If you specifically need F5-TTS, you are on route 1 or route 2. If you need a workable local voice for a real project and F5-TTS was only a means to that end, the app route removes the Python setup. Our no-Python local TTS on Mac guide walks through that path.

Here is the concrete path for a short cloning test on a Mac, using the MLX route. Prepare a clean reference clip first: ffmpeg -i source.wav -ac 1 -ar 24000 -sample_fmt s16 -t 10 ref.wav. Save the exact words spoken in that clip as your reference text. Then run python -m f5_tts_mlx.generate --text "Your script here." --ref-audio /path/ref.wav --ref-text "the transcript of ref.wav". Listen for dropped words and mispronunciations, because zero-shot cloning quality depends heavily on how clean and how representative your reference audio is. The same reference-quality rules apply to any cloning model, and our best local AI voice cloning apps for Mac roundup compares those workflows.

The license is the deciding factor for most Mac users

The repository states its position plainly: the code is MIT licensed, but the pretrained models are licensed under CC-BY-NC because of the Emilia training data. The Hugging Face model card confirms a cc-by-nc-4.0 license. NC means non-commercial, so personal projects, research, and internal experiments are the intended fit, and using the weights in a product or a paid client deliverable is a separate question that the license does not grant by default. That is not a reason to avoid F5-TTS for a portfolio piece or a study; it is a reason to read the terms before you monetize the output. Our commercial license guide for local TTS models explains how these restrictions differ across projects, and the safe habit is to check the exact model card for the checkpoint you downloaded.

F5-TTS compared with other local models on a Mac

Set F5-TTS beside the models a Mac creator is more likely to install through a packaged app, and the tradeoffs get clearer. The license signals below are the ones we track in our best local TTS models of 2026 comparison, but licenses change, so confirm each one on the model card before you commit.

Murmur ships all of the models in that table except F5-TTS, which is exactly why the license question matters. A creator who needs F5-TTS specifically is choosing a research-grade tool and taking on its terms; a creator who just needs a voice for a scripted video is usually better served by a model whose license already fits the work.

When F5-TTS is the right choice, and when it is not

• Choose F5-TTS if you are comfortable with a terminal, want the authors' original code, and are doing research, benchmarking, or a zero-shot cloning experiment.
• Choose it if the project is non-commercial, or if you have separately cleared the weights for the use you have in mind.
• Choose the MLX port if you want the simplest Apple silicon path of the two F5 routes and do not need the training tools.
• Skip F5-TTS if you need commercial rights out of the box; pick a model whose license permits your project and confirm it on the model card.
• Skip it if you want no Python at all, since a packaged Mac app gets you to finished narration faster.
• Skip it for live or real-time voice agents: F5-TTS is a generate-then-listen model, not a streaming service.
• Check memory before you start. Non-Kokoro local models generally want at least 16 GB of unified memory on an Apple Silicon Mac, and our 16 GB Mac guide covers what that means in practice.

Frequently asked questions