The whole training path is two actions: run one command that slices your recording, auto-labels it with Chinese ASR and writes the manifest; then open the WebUI and press train. Everything before that is a single clean WAV of twenty to thirty minutes.

Evidence note: this episode ships no workflow JSON — it is a training-mirror walkthrough, not a ComfyUI graph. All timings and settings below are the author's spoken values.

When this is useful

  • Building a reusable voice LoRA from your own recorded narration.
  • Replacing a manual slice-label-train pipeline with one command plus one button.
  • Projects where a multilingual TTS model with strong cloning is the requirement.
  • Batch-producing voiceovers in a consistent timbre.

How this differs from the previous build

The work was moved out of the notebook and into two surfaces.

  • Dataset prep is one command. Slicing, Chinese ASR labelling and manifest generation all happen in a terminal paste; the result is a manifest file in each folder.
  • Training is a button in the model's own WebUI. No custom training script, no config file to hand-edit.
  • Only one GPU class is offered. The author says special scripts are involved, so the mirror exposes 5090 only — which is also the fastest inference option.

Running it

  1. Prepare a clean WAV, twenty to thirty minutes long, with no background noise. The author exports from a video editor with the format set to WAV.
  2. Deploy the mirror and wait for it to self-start.
  3. Upload the audio in the dataset section and name it exactly as the notes specify.
  4. Copy the one-line preprocessing command from the description and paste it into a new terminal. It slices, labels and builds the manifest.
  5. Confirm each folder now holds a manifest file, then return to the WebUI.
  6. Click start training. The author's run completed 2000 steps quickly, saving a LoRA every 1000 steps, so two checkpoints.
  7. Collect the LoRA from the file manager path for VoxCPM output checkpoints.
  8. In the inference tab, refresh the model list, select the 2000-step LoRA, and generate.

Honest limits

  • The author's own result did not blow up, but the improvement over the base voice was not obvious — he suggests 1000-plus steps may already be enough.
  • English output is weaker than Chinese; the author's guess is that adding more English to the dataset would help, which he has not tested.
  • Dataset quality dominates everything: noisy audio produces a noisy clone regardless of step count.
  • The mirror only offers one GPU class, so you cannot trade speed for cost here.
  • No workflow JSON ships, and training hyperparameters are the mirror's, not exposed as a documented set.

Episode resources

Mirror (Compshare, referral code preserved)