FunCodec
modelscope/FunCodecReviewed by AI
Checked by RepoPicks Radar on
What is FunCodec?
FunCodec, from Alibaba's speech lab, is a research toolkit for neural audio codecs: it compresses speech into discrete codes at bitrates from 250 to 16000 bits per second and rebuilds the sound from them. Six pretrained models are published on Hugging Face and ModelScope, and the project also ships recipes for LauraTTS, a text-to-speech model that imitates a voice from a short sample. Everything runs through shell recipes on Linux with Python and one or more GPUs; there is no application, no interface and no release package.
How easy is it to install FunCodec?
E · developers only · Linux
This is for researchers on Linux: you clone the repository, run `pip install --editable ./`, then drive stage-based shell scripts such as `bash encoding_decoding.sh --stage 1 --gpu_devices "0,1"` with Kaldi-style `wav.scp` file lists. The setup.py declares Linux only, there is no graphical or single-command path, and training needs several GPUs.
Coming soon: install in 1 click with the RepoPicks app.
FunCodec on GitHub
- Stars
- 449on Nov 10, 2026
- Downloads
- —all releases
- Growth
- —no gain in the last reading
Stars over time
Facts
- Runs on
- Linux
- 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
- None
- Cost
- Free · MIT code and freely downloadable model weights; the cost is the GPU time.
- License
- MIT (Permissive)
- Last change on GitHub
- Jan 5, 2024
FunCodec risk signals (5)
- Attentionpackages installed over plain http
requirements.txt is a single line: `--editable ./ -i http://mirrors.aliyun.com/pypi/simple --trusted-host mirrors.aliyun.com`. It points pip at an unencrypted (http) package mirror and switches off the certificate check for it, so what gets installed depends on that mirror and on the network in between.
AI review
- Attentionno updates since january 2024
The last commit is from 2024-01-05 (1038 days before this check) and there are no releases at all, while the README still says "This project is still working on progress" and 26 issues are open. The dependency list asks for `torch>=1.12` and pins `typeguard==2.13.3`, which is the kind of pinning that breaks with current Python and PyTorch versions.
AI review
- Attentionimitates a voice from a sample
The repository topics include "voicecloning", and the README presents "LauraTTS... a powerful codec-based zero-shot text-to-speech synthesizer, which outperforms VALL-E in terms of semantic consistency and speaker similarity". Zero-shot means a short recording of a person is enough to make a new voice say anything, which can break voice and likeness rights.
AI review
- Infomodels downloaded from outside the repository
The weights are not in the repository: `bash encoding_decoding.sh --stage 0 --model_name ${model_name} --model_hub modelscope` (or `huggingface`) fetches them from alibaba-damo's pages on ModelScope or Hugging Face. Those model pages can carry their own terms, which are not the repository's MIT license.
AI review
- Infolicense statements do not match
LICENSE and setup.py's `license="The MIT License"` say MIT, but setup.py's classifiers list "License :: OSI Approved :: Apache Software License". The README adds that the project "contains various third-party components and some code modified from other repos under other open source licenses" (Kaldi, ESPnet, EnCodec).
AI review
Evidence is shown in its original language.
What Radar could not check about FunCodec
- Whether the install still completes with current Python and PyTorch versions after almost three years without commits
- How much disk and video memory the models need in practice, for encoding and for training
- Whether the pretrained models still download from ModelScope and Hugging Face
- Whether the quality claims against VALL-E hold outside the authors' own tests
- What the always-installed Alibaba Cloud storage client (oss2) and the training-time wandb package send out