rvc-free-colab — train RVC voice models free on Google Colab (2026)

Getting the old Rejekts "RVC Train" Colab notebook to actually train again on a current Colab runtime, without paying for anything. Patched for Python 3.12, numpy 2.x, torch 2.x and transformers 5.x — fairseq removed, runs on the free T4.

If this saved you an afternoon of dependency archaeology, please ⭐ star the repo — it genuinely helps others find it, and it's the only thing we ask in return. Issues and PRs welcome.

Keywords: RVC, Retrieval-based Voice Conversion, voice cloning, AI voice, Google Colab, free GPU, T4, fairseq alternative, ContentVec, HuBERT, train RVC model, Colab notebook fix.

I went to train a voice model with the well-known ardha27 / Rejekts RVC notebook and hit a wall: it was written back when Colab ran an older Python with numba 0.56, numpy 1.23 and fairseq. Colab has since moved to Python 3.12, numpy 2.x, torch 2.x and transformers 5.x, and the notebook now falls over in four different spots before it can finish a single epoch. None of the breakage is in the RVC training itself — it's all dependency rot and a couple of removed library APIs.

This repo is the small set of patches that fix it. It still runs on the free T4. You do not need Colab Pro or compute units.

How to use it

After the notebook's Install to Google Drive cell finishes, add one cell and run it:

!pip -q install av praat-parselmouth pyworld faiss-cpu transformers soundfile pyngrok
!wget -q https://raw.githubusercontent.com/webvijayi/rvc-free-colab/main/apply_fixes.py
!python apply_fixes.py /content/drive/MyDrive/project-main

Then run the rest in order: Preprocess, Extract Features, Train Index, Train Model. Skip the "Open the GUI" cell (more on that below). That's the whole change.

What was actually broken

Audio loading. The installer quietly dies at numba==0.56.4 (it won't build on 3.12), so av never gets installed and preprocessing throws ModuleNotFoundError: av. Install av and you hit the next thing: modern PyAV changed its API, so the notebook's av.open(file, 'rb') and add_stream(channels=1) calls fail too. Rather than chase the PyAV version, I rewrote load_audio() to shell out to ffmpeg, which is already on every Colab box and doesn't care about library churn.

Feature extraction. This one loads HuBERT through fairseq, and fairseq simply will not pip-install on Python 3.12 anymore (the omegaconf/hydra pins conflict and the build fails). The maintained RVC forks all dropped it, so I did the same: load the same weights through HuggingFace transformers using lengyue233/content-vec-best. For a v2 model you want hidden_states[12] (the 768-dim layer); v1 uses hidden_states[9] plus the final projection. One gotcha specific to the Rejekts notebook — it passes the args to extract_feature_print.py in a different order than upstream RVC, which made it read the experiment folder as "0", so the replacement figures out the path and the v1/v2 flag from the args directly instead of by position.

Training crashes after epoch 1. Losses log fine, then it dies in utils.plot_spectrogram_to_numpy when it tries to draw the TensorBoard preview. The culprit is np.fromstring(fig.canvas.tostring_rgb()) — np.fromstring is gone in numpy 2.x and tostring_rgb() was removed in matplotlib 3.8. Swapped both for np.frombuffer(fig.canvas.buffer_rgba()) and dropped the alpha channel.

Missing packages. pyworld and praat-parselmouth are also casualties of the aborted installer — they're in the pip line above.

One thing to know about Colab

Colab blocks the RVC Gradio web UI on the free tier. If you launch it you'll get "runtime disconnected — disallowed usage." That restriction is about the interactive UI and tunnel, not the training. The headless training cells run fine, so just don't open the GUI cell. Losses print to the cell and to logs/<model>/train.log, and TensorBoard over ngrok is optional. If you do get disconnected, reconnect and remount Drive — your project and progress are on Drive.

Heads up: training writes a fair amount to Drive (checkpoints, the spectrogram cache). If your Drive is near full you'll see truncated-file errors mid-training, so clear some space first or point the project at Colab's local disk.

Prefer no time limits? Train on Modal instead

If Colab's idle disconnects and session cap are getting in your way, there's a second free-GPU path: train on Modal. The job runs server-side so you can close your laptop, and because you control the Python version there, you pin 3.10 and stock RVC runs with no patches at all. Full app and a walkthrough of the Modal-specific gotchas are in MODAL.md.

What's in here

apply_fixes.py does all four patches in place against a downloaded project-main, and it's safe to re-run. The replacement extract_feature_print.py is embedded inside it. modal/train_rvc_modal.py + MODAL.md are the no-time-limit Modal path. There are no model weights and no audio in this repo — just the code and notes.

Credits

• The original notebook: ardha27/AI-Song-Cover-RVC
• RVC-Project/Retrieval-based-Voice-Conversion-WebUI
• HuBERT weights: lengyue233/content-vec-best
• transformers-based RVC references I leaned on: esnya/hf-rvc and Applio

Support this work

Built and maintained by webvijayi. It's free to use under MIT — the only condition is keeping the attribution. If it helped:

• ⭐ Star the repo so more people find it
• 🐛 Open an issue if a step breaks on a newer Colab runtime (they change things often)
• 🔁 Share it with anyone fighting the same fairseq / numpy 2.x errors

License

MIT — free for any use, including commercial, as long as the copyright/attribution to webvijayi is retained (that's all the MIT license requires). See LICENSE.