Beatrice Colab & Kaggle Trainer

Train your own Beatrice voice model for free, in your browser.
Two notebooks run the official Beatrice trainer on a free Google Colab or Kaggle GPU and give you a ZIP that the Beatrice voice changer imports in one click.

Open in Colab · Open in Kaggle · Website · Beatrice for Mac · Beatrice for Windows · Report a bug

---

✨ Why these notebooks

• 🆓 Free GPU, nothing to install. Colab's T4 or Kaggle's T4 x2 does the training. You only need a browser and a Google or Kaggle account.
• 🧠 The official trainer, pinned. Runs Project Beatrice's own trainer at the revision whose 2.0.0-rc.0 models the voice changer loads.
• 🔁 Survives disconnects. Checkpoints are saved to Google Drive (or /kaggle/working). Run the training step again and it carries on from the last save.
• 📦 One ZIP out. The export step builds beatrice_paraphernalia_<voice>_<step>.zip, which is what Upload Custom Model in Beatrice expects.
• 🎧 Full-quality audio. Your recordings are used at their original sample rate. Nothing is downsampled.
• 🗂️ One voice, one model. Each voice gets its own data and output folder, so voices and checkpoints never mix.

🧭 What's in a notebook

Step	Google Colab	Kaggle
1. Set up	Mounts Google Drive, installs the trainer	Checks Internet, installs the trainer
2a. Hugging Face dataset	Downloads a dataset's audio at full quality	Same
2b. Your recordings	Upload from your computer, or a ZIP/folder already in Drive	Copy from a Kaggle dataset you uploaded
3. TensorBoard (optional)	Listen to test conversions while it trains	Often blocked by Kaggle
4. Train	Trains one voice and resumes after a disconnect	Same
5. Export	Downloads the ZIP and keeps a copy in Drive	Download link plus the Output panel
6. Clean up (optional)	Deletes one voice, or everything, from Drive	Same, in /kaggle/working

_{There's no screenshot yet: the notebooks need a live Colab or Kaggle GPU session, so we haven't taken one.}

🚀 Get started

Google Colab

1. Click Open in Colab and sign in.
2. Choose Runtime → Change runtime type → T4 GPU and save.
3. Run Step 1 and allow access to Google Drive.
4. Run Step 2a to use a Hugging Face dataset, or Step 2b to upload your own recordings. Set speaker_name to the voice's name.
5. (Optional) Run Step 3 to open TensorBoard.
6. Run Step 4 and keep the tab open while it trains.
7. Run Step 5. Your browser downloads beatrice_paraphernalia_<voice>_<step>.zip, and a copy stays in MyDrive/Beatrice_Trainer/exports.

Kaggle

1. Click Open in Kaggle and sign in.
2. In the sidebar under Session options, set Accelerator to GPU T4 x2 and turn Internet on (this needs a phone-verified account). If you see Persistence, choose Files.
3. To use your own recordings, choose Add Input → Upload → New Dataset, then paste the dataset's path into Step 2b.
4. Run Steps 1, 2a or 2b, 4 and 5 in order. Each cell starts with the settings you can edit.
5. Download the ZIP from the link Step 5 prints, or open Output (/kaggle/working) in the sidebar, click ⋮ next to the file and choose Download.

Tip

Use clean recordings of one person: no music, no background noise, no reverb. Varied speech (questions, laughs, calm and loud lines) helps the model. WAV, FLAC, MP3, OGG/Opus and AIFF all work.

Important

Only train on voices you have the right to use: your own, or someone who agreed to it.

🎙️ Use your model in Beatrice

1. Open the Beatrice voice changer (Mac or Windows).
2. Open the model menu and choose Upload Custom Model.
3. Pick the ZIP from Step 5, type a name, choose a category and click Import. Beatrice looks up a portrait for the name and switches to the new voice.
4. Press the power button so it says LIVE, and talk.

Import the voice changer ZIP. The checkpoint (.pt.gz) and the whole-output ZIP are for training more later, and Beatrice can't import them.

⚙️ Training settings (Step 4)

Setting	Default	What it does
speaker_name	empty	Which voice to train. Empty means the voice from Step 2
training_steps	5000	Total steps (upstream default: 10000). To train longer, raise it and run Step 4 again
batch_size	8	Upstream uses about 9 GB of VRAM at this size, which fits a T4. Lower it if you run out of memory
learning_rate	0.00005	Same as upstream
save_every_n_steps	1000	How often a checkpoint and a model are saved, and test conversions are logged
warmup_steps	1000	Learning-rate warm-up (upstream: 5000)
num_workers	0	Data-loading processes. 0 means one per CPU core
resume_from_checkpoint	on	Off moves the previous run to output/<voice>_old_<time> and starts fresh
log_loss_curves	off	Adds loss curves to TensorBoard. The upstream trainer leaves them off on purpose: they don't show quality

Where are my files?

	Google Colab	Kaggle
Recordings	MyDrive/Beatrice_Trainer/data/<voice>	/kaggle/working/data/<voice>
Checkpoints and models	MyDrive/Beatrice_Trainer/output/<voice>	/kaggle/working/output/<voice>
Exported ZIPs	MyDrive/Beatrice_Trainer/exports	/kaggle/working

Each saved step creates output/<voice>/paraphernalia_<voice>_<step>/ (the model) and checkpoint_<voice>_<step>.pt.gz (for resuming). Checkpoints are large, so plan for 1–2 GB per voice.

🧩 How it works

flowchart LR
  hf([Hugging Face dataset]) --> data
  rec([Your recordings]) --> data
  data["Step 2<br/>data/voice"] --> train["Step 4<br/>official Beatrice trainer<br/>2.0.0-rc.0"]
  train --> out["output/voice/<br/>paraphernalia_voice_step"]
  out --> zip["Step 5<br/>beatrice_paraphernalia_voice_step.zip"]
  zip --> app([Beatrice<br/>Upload Custom Model])

Loading

Step 1 makes sure the runtime has what the trainer needs: torchaudio 2.1 to 2.8 (it installs torch and torchaudio 2.8.0 if the runtime has something newer), pyworld, soundfile and tensorboard. It then clones fierce-cats/beatrice-trainer at commit f34836d with its pretrained files.

🛠️ Troubleshooting

Step 1 reinstalls PyTorch. Is that normal?

Yes, if the runtime ships torchaudio 2.9 or newer. The Beatrice trainer needs torchaudio older than 2.9, so Step 1 installs torch 2.8.0 and torchaudio 2.8.0 (CUDA 12.6). That takes a few minutes, and only happens once per session. pyworld is compiled from source, which also takes a minute.

"No GPU found" or training is extremely slow

On Colab choose Runtime → Change runtime type → T4 GPU. On Kaggle set Accelerator to GPU T4 x2. Then run Step 1 again. On Kaggle, if training fails with no kernel image is available, the GPU is too old for this PyTorch build: switch from P100 to T4.

Colab disconnected while training

Reconnect, run Step 1, then Step 4 with resume_from_checkpoint on. Training continues from the last save, so you lose at most save_every_n_steps steps. If it stops straight away, the checkpoint has already reached training_steps: raise it to train more.

Beatrice says required model files are missing

You picked the checkpoint or the whole-output ZIP. Run Step 5 with Voice changer ZIP and import beatrice_paraphernalia_<voice>_<step>.zip. It must contain speaker_embeddings.bin and waveform_generator.bin, plus the other model files.

Kaggle: installs or downloads fail

Turn Internet on in the sidebar. Kaggle only offers it to accounts with a verified phone number.

I trained with an older version of this notebook

Older versions saved every run straight into output/. Its resume check looked in the wrong folder, so a second run stopped with checkpoint_latest.pt.gz already exists. To continue an old run, move the files in output/ into output/<voice>/ and run Step 4. Older versions also resampled Hugging Face datasets to 16 kHz, which makes the voice sound muffled, so run Step 2a again for full-quality audio.

Can I use RVC models or other voice files?

No. RVC models are a different kind of model and can't be converted. Train a Beatrice voice here, or grab a ready-made one from Beatrice-voice-models.

🌐 The Beatrice family

🍎 Beatrice for Mac	The real-time voice changer for Apple Silicon Macs
🪟 Beatrice for Windows	The same app for Windows 10 and 11
🌍 Website	Downloads, guides and voice samples
🧠 Trainer for Mac · Windows	Train on your own computer
📂 Dataset Web UI	Prepare recordings for training
🎭 Voice models	Ready-made models to import

🙏 Credits

• Beatrice trainer and pretrained models by Project Beatrice (fierce-cats/beatrice-trainer, MIT)
• Beatrice engine by prj-beatrice
• The notebook flow started from w-okada/beatrice-trainer-colab
• Free GPUs from Google Colab and Kaggle

📜 License

The notebooks are MIT licensed, © 2026 Project Beatrice V2 (see LICENSE). The Beatrice trainer and its pretrained models are MIT licensed by Project Beatrice. Datasets you train on keep their own licences, and a voice model doesn't give you rights to anyone's voice.

If you trained a voice you love, ⭐ star the repo. It really helps.

Back to top ↑