What did the VoxCPM 1.5 LoRA run actually improve?
By Wei Jie Chee · 07 Feb 2026, 00:00 Z
Download printable cheat-sheet (CC-BY 4.0)The run establishes a reproducible adaptation path, while checkpoint listening must decide whether voice identity improved without adding instability.
Do this first
Compare the base model and preserved checkpoints on the same held-out prompts before choosing by training loss.
Keep this boundary in view
One English speaker, corpus, adapter configuration, and checkpoint set do not establish general VoxCPM adaptation quality.
Inspect the evidence and what would change the answer
First-party LoRA training and checkpoint evidence
What supports it
The article preserves the dataset, training route, checkpoint outputs, and evaluation conditions for one voice.
Change our mind if
Matched listening shows no repeatable identity gain, or a different adapter setting improves quality without the observed trade-offs.
Claim TTS-VOXCPM15-LORA-2026-08 / v1, checked 7 Aug 2026
60-second takeaway
VoxCPM produced a strong practical result in this benchmark once we stabilized dataset prep and used a clean inference protocol.
The best balance of quality and stability came fromstep_0004000in our run.
Prompted inference can over-copy prompt noise, so prompt quality matters as much as checkpoint choice.
Where this fits
- For founders: VoxCPM is a viable production candidate from this benchmark.
- For engineers: use this page for train recipe, checkpoint pick logic, and inference defaults.
For the series overview matrix, see:
Experiment setup
- Base model: VoxCPM1.5
- Dataset: IMDA NSC
FEMALE_01 - Audio prep: resampled to 44.1 kHz for VoxCPM1.5 path
- Hardware: RTX 3090 Ti 24 GB
- Training mode: LoRA fine-tuning
Best checkpoint logic
We tracked validation total loss across steps and selected the strongest zone by both trend and listening:
- Best recorded validation total in this run was at
step_0004000. - Later checkpoints remained usable, but were not consistently better on subjective naturalness.
Audio evidence
Best practical sample (this run)
Settings: no-prompt, no denoiser, long text test.
Failure modes we saw
- Prompted outputs can inherit prompt-room noise strongly.
- Denoisers can clean hiss but also shift timbre and bandwidth perception.
- Long-form outputs are sensitive to prompt clip quality and consistency.
Recommended inference settings
For this exact benchmark setup:
- Start from
step_0004000as default checkpoint. - Use no-prompt generation first to estimate model prior naturalness.
- Add prompt only when you need stronger speaker lock.
- Use denoiser only when hiss/noise is clearly audible.