FireRedTTS3
FireRedTeam/FireRedTTS3Reviewed by AI
Checked by RepoPicks Radar on
What is FireRedTTS3?
FireRedTTS3 is the published code and model for a speech system from a Chinese research team. Give it a few seconds of someone speaking and it reads your text in that voice, across 24 languages and 21 Chinese dialects. A second model invents a voice from a written description ("gentle young woman, slightly slow, a bit playful") and edits existing recordings: swap, add or remove words, or change speed, pitch and volume. It is Python code with a small Gradio web page, aimed at people comfortable with machine-learning setups and a graphics card.
FireRedTTS3 is free and open source (Apache-2.0) and can replace ElevenLabs, Fish Audio, and Descript.
How easy is it to install FireRedTTS3?
D · terminal, several tools or a specific graphics card · Windows
- Windows D
- Linux D
A D on both: you clone the repository, "pip install -r requirements.txt" (which pins flash_attn 2.8.3, a package that normally wants an NVIDIA card and a compiler), download the weights with the Hugging Face or ModelScope command line, optionally fetch a language-detection file with curl, and then either write Python or start the Gradio page with "python gradio_base.py --host 0.0.0.0 --port 7860".
Coming soon: install in 1 click with the RepoPicks app.
FireRedTTS3 on GitHub
- Stars
- 1,720on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +2stars in 6 days
Stars over time
Facts
- Runs on
- Linux, Windows
- You need to install
- Python, Graphics card
- To build from source
- Python, Graphics card
- Disk space
- 950 MB–4.3 GB · estimate
- Accounts or keys
- Other (LLM_TN_API_KEY)
- Cost
- Free · Code and model weights are free (Apache-2.0 code). You pay in hardware or rented GPU time. The optional whole-language text normalisation calls a paid AI endpoint of your choice.
- License
- Apache-2.0 (Permissive)
- Can replace
- ElevenLabs, Fish Audio, Descript
- Last change on GitHub
- Sep 8, 2026
FireRedTTS3 risk signals (5)
- AttentionDependencies with known vulnerabilities
Pinned dependencies with known vulnerabilities (OSV): torch: GHSA-63cw-57p8-fm3p, GHSA-qfhq-4f3w-5fph, GHSA-rrmf-rvhw-rf47; flash_attn: GHSA-7g5w-pq96-8c5w, PYSEC-2026-2477; transformers: GHSA-x9r9-c232-4q39, GHSA-xrqw-3rrv-vx5w, PYSEC-2026-3929
automatic check
- Attentionvoice cloning and putting words in recordings
The model clones a voice from a short reference clip in 24 languages and 21 Chinese dialects, and the Instruct model edits real audio at the word level: "Semantic Edit ... Content-level editing: insertion / deletion / substitution by instruction", with the example "Replace 'cats' with 'dogs'." That means making a recording of a real person say something they never said. The README mentions no watermark on the generated audio.
AI review
- Attentionauthors restrict use to research
Usage Disclaimer: "The project incorporates zero-shot voice cloning functionality; Please note that this capability is intended solely for academic research purposes", "DO NOT use this model for ANY illegal activities" and "The developers assume no liability for any misuse of this model." The code license itself is Apache-2.0, so the stated intent is narrower than the licence.
AI review
- Infomodel files come from outside github
The weights are pulled from Hugging Face ("hf download FireRedTeam/FireRedTTS3 --local-dir pretrained_models/") or ModelScope, and the optional language detector is fetched with "curl -L -o fireredtts3/utils/llm_tn/models/lid.176.ftz https://dl.fbaipublicfiles.com/fasttext/...". None of this is in the repository and the sizes are not stated.
AI review
- Infogradio demo binds to all interfaces
The documented command is "python gradio_base.py --host 0.0.0.0 --port 7860", which answers on every network interface with no password, so on a shared or office network other machines can reach the voice-cloning page.
AI review
Evidence is shown in its original language.
What Radar could not check about FireRedTTS3
- Whether the pinned packages (torch 2.8.0, flash_attn 2.8.3, transformers 5.6.2) install on your machine, and on which cards
- How much video memory and disk the two models need, which the README never says (the 950-4450 MB figure is an estimate)
- Whether the quality tables, which were produced by the authors and partly scored by Gemini, hold on your own recordings
- Whether generated or edited audio carries any mark showing it is synthetic
- What the optional LLM text normalisation sends to the endpoint you configure, and what that costs