AI models

F5-TTS

Decision matrix

F5-TTS model decision matrix
Overview

F5-TTS is an open-source (MIT code, CC-BY-NC pre-trained weights) text-to-speech model you run on your own GPU, generating fluent speech via diffusion transformer and flow matching, with sub-300ms latency reported on a single high-end GPU. Pick it when you want speech generation with no per-character API cost and are willing to run your own hardware; the pre-trained weights aren't licensed for commercial use.

Other good fits
  • Non-commercial or research projects, where the CC-BY-NC license on the pre-trained weights isn't a blocker.
Pros
Read pros

F5-TTS is an open-source text-to-speech model using a Diffusion Transformer with ConvNeXt V2 and flow matching to generate fluent, natural-sounding speech from text. The code is MIT-licensed and free; you run it yourself, with no hosted API and no per-character billing. It supports NVIDIA (CUDA 12.4+), AMD (ROCm) and Intel GPUs, with CPU inference possible on Apple Silicon.

Pick it when…

  • You want speech generation with zero per-character API cost, at the price of running your own GPU.
  • Your project is non-commercial or research-oriented, matching the CC-BY-NC license on the provided pre-trained weights.
  • You want low-latency local generation; the project reports 253ms latency on a single high-end GPU (an NVIDIA L20).
Cons
Read cons

The CC-BY-NC license on the pre-trained weights is a real constraint: it means the ready-to-use model isn't cleared for commercial use, only the code itself is MIT-licensed. Local generation needs a real GPU for good performance; CPU-only inference on less capable hardware will be noticeably slower than the benchmark figures suggest.

Pick something else when…

  • You need the output cleared for commercial use without training your own weights. Cartesia or ElevenLabs include commercial licensing on their paid tiers.
  • You don't want to manage GPU hardware or local inference at all. A hosted API removes that setup entirely, at a per-character or per-credit cost.
  • You want speech generation integrated with a broader local AI workspace. AnythingLLM or a similar local-first tool may fit better if TTS is one feature among several you need.

Avoid if

  • You don't have a capable GPU and don't want to manage local inference; a hosted API like Cartesia or ElevenLabs is far less setup.
Context

Not documented yet.

Requirements

F5-TTS needs Python 3.10+ and FFmpeg, plus a GPU: NVIDIA with CUDA 12.4+, AMD with ROCm 6.2/7.2, or Intel with XPU support; Apple Silicon can run it on CPU, more slowly. Install via pip for straightforward use, or clone the GitHub repository if you want to train or fine-tune the model yourself.

Licence & cost

The code is free (MIT); running it costs only your own hardware and electricity, no per-character or per-minute API fee. The pre-trained model weights carry a separate CC-BY-NC (non-commercial) license because of their training data (the Emilia in-the-wild dataset), so check that license carefully before using the provided weights in anything commercial; training or fine-tuning your own weights on your own data would carry different licensing considerations.

First task

Documentation-based suggestion — not run here.

Install F5-TTS locally and generate a short, made-up announcement using a stock or sample voice, then listen to the output against the text for pronunciation and naturalness, the same way you'd evaluate any text-to-speech system.

Done when you've generated and listened to at least one clip, and can say whether local generation quality and speed on your hardware are good enough for your actual project.

Links & evidence
Documentation-based · not hands-on · Sep 23, 2026

Rests on F5-TTS's GitHub repository, read on 2026-09-23. No model was installed or run for this review.

  • F5-TTS GitHub repository · Sep 23, 2026 — MIT code license, CC-BY-NC pre-trained weights, GPU requirements (NVIDIA/AMD/Intel), 253ms latency benchmark on an L20 GPU.
  • Cartesia pricing · Sep 23, 2026 — Comparison alternative with hosted, commercially licensed text-to-speech.
  • ElevenLabs pricing · Sep 23, 2026 — Comparison alternative with hosted, commercially licensed text-to-speech and voice cloning.
Further review notes & alternatives

Alternatives and trade-offs

  • Cartesia wins when you need commercial licensing and hosted convenience over local, non-commercial generation.
  • ElevenLabs wins for the same reason, with a broader toolkit beyond just speech generation.
  • Deepgram wins when speech recognition, not generation, is the actual need.

Learn more

Catalogue facts

Generate expressive speech from text using an open-source TTS system. Install it with pip or Docker on a PyTorch-capable machine. Use pretrained models for local speech experiments and drafts.

free · self-host

  • text-to-speech
  • voice cloning
  • open weights
  • local inference
  • non-commercial licence
  • open-source

Browse all tools