{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# Train a voice character on Colab\n", "\n", "Reusable for any character - mirrors `studio/voice.py`'s `train()` (same Applio\n", "commit, same 4 stages, same flags), so the model works with both Swara Studio\n", "(desktop) and the Swara Sound web app's Voices page.\n", "\n", "**Before you start:** Runtime -> Change runtime type -> **T4 GPU**.\n", "\n", "**1. Get your recordings into a zip.** Either:\n", "- **Swara Studio (desktop):** Voices -> fill in Name + Recordings -> **Prepare\n", " for Colab...**. Writes `_dataset.zip` next to the app.\n", "- **Swara Sound (web):** Voices -> name it, pick your recordings, **Zip &\n", " upload dataset to Drive**. This uploads the zip straight to\n", " `My Drive/SwaraStudioVoices/_dataset.zip` for you - skip to step 4.\n", "\n", "**2. Set `MODEL_NAME`** in the next cell to that same name (letters/numbers,\n", "matches what you typed).\n", "\n", "**3. Give the notebook the zip - two ways, pick one:**\n", "- **Upload it directly** (simplest, nothing to set up). When the \"Get the\n", " dataset\" cell runs, an upload button appears - pick the `_dataset.zip`\n", " from your computer. Fine for datasets under a few hundred MB.\n", "- **Or put it on Google Drive first**, at exactly:\n", " `My Drive/SwaraStudioVoices/_dataset.zip`\n", " The notebook checks that exact path and uses it automatically, skipping the\n", " upload prompt - handy if the notebook disconnects and you need to re-run.\n", " (The web app's uploader already puts it exactly here.)\n", "\n", "**4. Run every cell top to bottom.** Keep the tab open (roughly 1-3 hours\n", "depending on epochs below - don't touch the tab while the training cell runs).\n", "\n", "**5. The last cell packages checkpoints at epochs 150 / 200 / 250** (so you can\n", "A/B them), copies them to `My Drive/SwaraStudioVoices/_model/`, and also\n", "offers them as a direct browser download.\n", "\n", "**6. Bring the files back.** You need `.index` and **one**\n", "`_e.pth` (start with `150e` - least risk of over-training artefacts on\n", "a small dataset). Either:\n", "- **Desktop:** put both in `C:\\SwaraStudio\\engines\\rvc\\applio\\logs\\\\`\n", " and rename the `.pth` to `.pth`. Voices page then lists ``.\n", "- **Web:** Voices -> **Import a trained voice** -> name it, pick the two\n", " downloaded files (or grab them from Drive), **Import**.\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# --- settings ---------------------------------------------------------------\n", "MODEL_NAME = \"MyVoice\" # <-- must match the Name you used in Swara Studio\n", "SAMPLE_RATE = 40000 # Applio trains at 32000 / 40000 / 48000\n", "EPOCHS = 250 # ~10-15 min of clean audio -> 150-250 is the useful band\n", "BATCH_SIZE = 8 # T4 16 GB handles 8-12 at 40k\n", "SAVE_EVERY = 25 # checkpoint every N epochs\n", "DRIVE_ROOT = \"/content/drive/MyDrive/SwaraStudioVoices\" # one folder for every character\n", "DATASET_ZIP = f\"{MODEL_NAME}_dataset.zip\" # exactly what \"Prepare for Colab\" writes\n", "APPLIO_COMMIT = \"69b298e6336509b4caf01e3c2e43892b9075eb2a\" # == engines/rvc/applio\n", "\n", "!nvidia-smi -L" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from google.colab import drive\n", "drive.mount('/content/drive')" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# --- install Applio, pinned to the same commit the app uses ----------------\n", "%cd /content\n", "![ -d Applio ] || git clone https://github.com/IAHispano/Applio.git\n", "%cd /content/Applio\n", "!git checkout {APPLIO_COMMIT}\n", "!pip install -q -r requirements.txt\n", "!python core.py prerequisites\n", "# some clones lack this optional file; the model-extraction step hard-crashes\n", "# without it even though it only holds an author name\n", "import os, json\n", "os.makedirs('assets', exist_ok=True)\n", "if not os.path.exists('assets/config.json') or os.path.getsize('assets/config.json') == 0:\n", " json.dump({'model_author': MODEL_NAME}, open('assets/config.json', 'w'))\n", "print('assets/config.json ok')" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# --- get the dataset zip: Drive first, else ask you to upload it -----------\n", "import os, shutil\n", "from pathlib import Path\n", "\n", "os.makedirs(DRIVE_ROOT, exist_ok=True)\n", "drive_zip = Path(DRIVE_ROOT) / DATASET_ZIP\n", "local_zip = Path(\"/content\") / DATASET_ZIP\n", "\n", "if drive_zip.exists():\n", " print(f\"Found it on Drive: {drive_zip}\")\n", " shutil.copy(drive_zip, local_zip)\n", "else:\n", " print(f\"Not on Drive at {drive_zip}.\")\n", " print(\"Upload the dataset zip now (Swara Studio -> Voices -> Prepare for Colab):\")\n", " from google.colab import files\n", " uploaded = files.upload()\n", " assert uploaded, \"no file uploaded - re-run this cell and pick the zip\"\n", " fn = next(iter(uploaded))\n", " shutil.move(fn, local_zip)\n", "\n", "print(f\"Using dataset: {local_zip} ({local_zip.stat().st_size / 1e6:.1f} MB)\")" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# --- unzip into Applio's dataset folder -------------------------------------\n", "import zipfile, glob\n", "dst = f\"/content/Applio/dataset/{MODEL_NAME}\"\n", "os.makedirs(dst, exist_ok=True)\n", "with zipfile.ZipFile(local_zip) as z:\n", " z.extractall(dst)\n", "exts = (\".wav\", \".mp3\", \".flac\", \".ogg\", \".m4a\")\n", "n = sum(1 for f in glob.glob(f\"{dst}/**/*\", recursive=True) if f.lower().endswith(exts))\n", "assert n > 0, f\"No audio found after extracting {local_zip} - is it the right zip?\"\n", "print(f\"{n} audio file(s) ready in {dst}\")" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# --- Stage 1/4: preprocess ----------------------------------------------\n", "%cd /content/Applio\n", "!python core.py preprocess --model-name {MODEL_NAME} \\\n", " --dataset-path /content/Applio/dataset/{MODEL_NAME} \\\n", " --sample-rate {SAMPLE_RATE} --cut-preprocess Automatic \\\n", " --noise-reduction --normalization-mode post" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# --- Stage 2/4: extract features (pitch + speaker embedding) ------------\n", "!python core.py extract --model-name {MODEL_NAME} --sample-rate {SAMPLE_RATE} \\\n", " --f0-method rmvpe --gpu 0 --embedder-model contentvec" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# --- Stage 3/4: train (the long one - keep the tab open, don't touch it) --\n", "!python core.py train --model-name {MODEL_NAME} --total-epoch {EPOCHS} \\\n", " --sample-rate {SAMPLE_RATE} --batch-size {BATCH_SIZE} --gpu 0 \\\n", " --save-every-epoch {SAVE_EVERY} --save-every-weights \\\n", " --pretrained --vocoder HiFi-GAN" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# --- Stage 4/4: build the .index retrieval file ----------------------\n", "!python core.py index --model-name {MODEL_NAME} --index-algorithm Auto" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# --- package usable models, copy to Drive, offer a direct download ---------\n", "# Applio's own end-of-training extraction is flaky, so do it here from the\n", "# G_ checkpoints. Packages epochs 150 / 200 / 250 so you can A/B.\n", "%cd /content/Applio\n", "import os, json, glob, shutil, torch\n", "os.makedirs('assets', exist_ok=True)\n", "if not os.path.exists('assets/config.json') or os.path.getsize('assets/config.json') == 0:\n", " json.dump({'model_author': MODEL_NAME}, open('assets/config.json', 'w'))\n", "from rvc.train.utils import HParams\n", "from rvc.train.process.extract_model import extract_model\n", "\n", "EXP = f\"logs/{MODEL_NAME}\"\n", "cfg = HParams(**json.load(open(f\"{EXP}/config.json\")))\n", "\n", "def step_for(epoch):\n", " gs = sorted(int(os.path.basename(p)[2:-4]) for p in glob.glob(f\"{EXP}/G_*.pth\"))\n", " per = gs[0] // SAVE_EVERY if gs else 31 # steps per epoch\n", " want = epoch * per\n", " return min(gs, key=lambda s: abs(s - want)) if gs else None\n", "\n", "made = []\n", "for epoch in (150, 200, 250):\n", " step = step_for(epoch)\n", " g = f\"{EXP}/G_{step}.pth\" if step else None\n", " if not g or not os.path.exists(g):\n", " print(\"skip (no checkpoint near epoch)\", epoch); continue\n", " ck = torch.load(g, map_location=\"cpu\", weights_only=False)\n", " sd = ck[\"model\"] if \"model\" in ck else ck\n", " out = f\"{EXP}/{MODEL_NAME}_{epoch}e.pth\"\n", " extract_model(ckpt=sd, sr=cfg.data.sample_rate, name=MODEL_NAME,\n", " model_path=out, epoch=epoch, step=step, hps=cfg, vocoder=\"HiFi-GAN\")\n", " if os.path.exists(out):\n", " made.append(out); print(\" OK\", out)\n", "\n", "idx = glob.glob(f\"{EXP}/*.index\")\n", "assert made and idx, f\"nothing to package: pth={made} idx={idx}\"\n", "\n", "drive_dst = f\"{DRIVE_ROOT}/{MODEL_NAME}_model\"\n", "os.makedirs(drive_dst, exist_ok=True)\n", "for p in made + idx:\n", " shutil.copy(p, drive_dst); print(\" copied to Drive:\", os.path.basename(p))\n", "\n", "print(f\"\\nSaved to Drive: {drive_dst}/\")\n", "print(f\"Also downloading {os.path.basename(made[0])} + {os.path.basename(idx[0])} to your browser...\")\n", "from google.colab import files\n", "files.download(made[0])\n", "files.download(idx[0])\n", "\n", "print(f\"\\nOn the laptop: put {os.path.basename(idx[0])} and ONE *_e.pth (start with 150e) in\")\n", "print(f\"C:\\\\SwaraStudio\\\\engines\\\\rvc\\\\applio\\\\logs\\\\{MODEL_NAME}\\\\ and rename the pth to {MODEL_NAME}.pth\")" ] } ], "metadata": { "accelerator": "GPU", "colab": { "provenance": [], "gpuType": "T4" }, "kernelspec": { "display_name": "Python 3", "name": "python3" }, "language_info": { "name": "python" } }, "nbformat": 4, "nbformat_minor": 0 }