Essential cookies keep Instavar working. Optional analytics help us understand how the site is used and link your first-visit source to your Studio account after sign-in, for up to 180 days. Cookie Policy
Manage Cookie Preferences
Service reliability telemetry, including Sentry error monitoring and Vercel Speed Insights, stays enabled so we can secure the product and diagnose failures.
Landscape scannerCompress the field into models worth a closer look.
A useful shortlist keeps only models that match the intended voice, adaptation route, hardware, latency, and licence.
Do this first
Choose the two constraints you cannot relax, then remove models that fail either one before comparing demos.
Keep this boundary in view
This is a shortlist for further testing, not a universal quality ranking or proof on your speaker and prompts.
Inspect the evidence and what would change the answer
Mixed first-party runs and model-specific research
What supports it
Instavar has trained or run several shortlisted models, but evidence depth differs by model and experiment.
Change our mind if
A currently omitted model repeatedly meets the same constraints and avoids the costly failures on a representative local test.
Claim TTS-FIELD-2026-08 / v1, checked 7 Aug 2026
60-second takeaway We ran a consistent FEMALE_01 benchmark across open-source TTS models on an RTX 3090 Ti (24GB). VoxCPM 1.5 and Qwen3-TTS 1.7B produced deployable LoRA outputs. IndexTTS2 step 14,000 is the current production narration default after it ranked first in a later same-transcript comparison with VoxCPM 2 full SFT step 2,000 and VoxCPM 1.5 full SFT step 2,999. CosyVoice3's corrected LoRA run completed, but later matched checks failed the content gate. For the lowest-friction LoRA experiment, start with VoxCPM 1.5 or Qwen3-TTS. For the current tested production default, use the pinned IndexTTS2 step 14,000 path and review its custom model license for your intended use.
Evidence updated: August 17, 2026. Repository status and visibility were checked against the current registered heads.
What this benchmark covers
This is a practitioner-oriented comparison, not an academic leaderboard. We evaluated four models under the same conditions:
Dataset: IMDA NSC FEMALE_01 - a TTS corpus slice with a posh or UK-influenced Singaporean voice profile
Hardware: one NVIDIA RTX 3090 Ti (24 GB VRAM)
Goal: produce voice-cloned audio suitable for AI-generated video narration (A-roll use case)
Evaluation: qualitative listening on naturalness, long-text stability, accent retention, and operational friction
We are not measuring WER or MOS scores from automated tools. We are measuring whether the output sounds production-ready to a human listener on a video platform.
The official NSC structure supports the FEMALE_01 provenance. The directory
label should not be treated as independent proof of the number or identity of
people represented in every local artifact.
The seven registered repositories behind this work
The benchmark results are only one layer of the work. We also maintain six
model-specific repositories and one shared evaluator. Together they cover
dataset lineage, LoRA and full-SFT lifecycles, checkpoint reload, runtime
qualification, content-faithfulness checks, packaging, restore evidence, and
blind-listening preparation.
Six repositories are public. The Instavar VoxCPM companion is currently private,
so public readers can inspect the upstream VoxCPM repository
and our on-site findings, but not the private companion's current implementation
evidence. The broader /open-source index therefore lists only repositories
GitHub currently exposes publicly.
These repositories do not all have the same proof level. Supported means the
named path has executable evidence for its stated scope. Experimental means a
bounded smoke or narrow qualification exists. Negative result
means the path
ran but failed its content or evaluation gate. An inherited runtime is not
treated as adapted-artifact support until the adapted artifact passes that
runtime's conformance checks.
The 100-step LoRA did not establish an overall adaptation benefit; optimized runtimes remain experimental and blind ratings are pending
The current repository ledger reports 16 of 16 implementation families with a
canonical owner and reachable evidence. That is an ownership and traceability
result, not a claim that every runtime or voice is production-ready.
The model families in this benchmark
VoxCPM 1.5
VoxCPM 1.5 uses a LoRA finetuning path that fits within 24GB VRAM without modification. Training is straightforward with standard train/val splits.
Dimension
Result
Finetuning approach
LoRA
Best checkpoint (this run)
step_0004000
Long-text stability
Good
Prompt sensitivity
Moderate - use clean prompt clips
Production-ready?
Yes
Key insight: No-prompt generation at step 4000 gave the best naturalness. Prompted inference copied prompt room noise into the output, which was audible on studio playback. Use prompt only when strong speaker lock is required.
VoxCPM 2
VoxCPM 2 is the newer full-SFT evidence point. LoRA fits comfortably, and full SFT now fits on an RTX 3090 Ti when the run uses gradient checkpointing, paged 8-bit optimizer state, allocator tuning, and a clean manifest.
Dimension
Result
Finetuning approach
LoRA or full SFT
Best checkpoint evidence
Full SFT selected step 2000 by held-out validation in the split run
Memory stack, dataset cleanup, disk, and validation
Production-ready?
Promising - validation-selected checkpoint available
Key insight: Full SFT on a 24GB consumer GPU is no longer just theoretical, but it is not the low-friction starting point. Use VoxCPM 1.5 LoRA when you need a fast deployable path; use VoxCPM 2 full SFT when you are testing whether deeper adaptation beats adapter speed.
Qwen3-TTS 1.7B
Qwen3-TTS 1.7B with LoRA was the model where adapter scale mattered most. Scale 1.0 over-steered and produced noisy outputs; scale 0.3 to 0.35 sounded stable.
Dimension
Result
Finetuning approach
LoRA
Best checkpoint (this run)
Epoch 10
Best LoRA scale
0.3 to 0.35
Long-text stability
Good with SDPA backend
Prompt sensitivity
Low - robust to formatting variation
Production-ready?
Yes
Key insight: Adapter scale and checkpoint selection interact. Run the same held-out samples across a bounded scale sweep before committing to a checkpoint. Scale 1.0 over-steered this checked voice; that does not prove it is wrong for every dataset or adapter.
IndexTTS2
IndexTTS2 uses full SFT (not LoRA). It requires more careful checkpoint management because the training loop had crash recovery issues in our run.
Dimension
Result
Finetuning approach
Full SFT
Best checkpoint (this run)
model_step14000.pth
Long-text stability
Good
Crash recovery
Required explicit resume management
Production-ready?
Yes - with operational caution
Key insight: Keep ALL checkpoints until you've done a listening eval sweep. The retention policy deleted older checkpoints before we could test them. Pin the best checkpoint explicitly once identified - don't rely on automatic deletion logic.
CosyVoice3
CosyVoice3 was the outlier. Our first full-SFT run did not reach production quality, while the corrected LoRA rerun was much more stable.
Dimension
Result
Finetuning approach
LoRA for the corrected run
Run status
Training and reload validated; matched content result failed
The later same-conditioning checks produced valid audio files, but Base and the
epoch-12 adapter both showed semantic corruption, high WER, and repetition. A
corrected instruction route also failed the requested-text gate. CosyVoice2
zero-shot remains useful as a control reference, but a valid WAV or successful
reload is not enough to promote the adapted artifact. See CosyVoice LoRA Fine-Tuning: What Worked, What Didn't, and the Rerun Plan for the full diagnostics and rerun plan.
Use IndexTTS2 step 14,000 for the FEMALE_01 narration workflow. In the May
10 same-transcript review, it ranked first for naturalness, target-profile fit,
and the desired accent balance against VoxCPM 2 and VoxCPM 1.5 full-SFT
checkpoints. This is a result for that voice profile, transcript set, and review,
not a universal model ranking.
If you need deployable output fastest
Start with VoxCPM 1.5 step 4000. It had the lowest setup friction and the cleanest no-prompt output in our run. LoRA training is straightforward and the checkpoint selection rule is simple.
If you need LoRA-style adapter control
Use Qwen3-TTS 1.7B LoRA. The scale parameter gives you a post-training knob to tune output strength without retraining. This is valuable when you want to fine-tune the output on different content types without full retraining cycles.
If you need the most reproducible full-SFT baseline
Use IndexTTS2 if you want the established full-SFT baseline. Use VoxCPM 2 full SFT if you specifically want to test the newer consumer-GPU full-SFT path. It is feasible on 24GB, but only with gradient checkpointing, paged optimizer state, clean manifests, and post-hoc validation.
If you want to evaluate CosyVoice
Use CosyVoice2 as a zero-shot baseline. Do not deploy the failed CosyVoice3
full-SFT run or promote the current LoRA artifact: the matched checks produced
valid WAVs but failed requested-text and repetition gates. Any frontend or
instruction-route correction needs a new frozen evaluation plan before another
listening decision.
What is IMDA NSC FEMALE_01?
IMDA NSC is the National Speech Corpus published by Singapore's Infocomm Media Development Authority. FEMALE_01 is a named TTS voice-profile slice with a posh or UK-influenced Singaporean sound. We use it because that profile makes differences in naturalness, target closeness, and accent drift easy to hear. Performance on this profile does not establish performance on other Singaporean or non-American-English voices.
Audio evidence
All audio samples from this benchmark are published in the individual model deep dives. Listen to them side by side before making a deployment decision.
Can I run these models on a single RTX 3090 Ti (24GB)?
Specific recipes in this benchmark completed training or inference within 24GB, but that does not mean every adaptation mode or sequence length fits. The full feasibility notes - including peak VRAM, runtime, and recipe availability - are covered in Voice Cloning on a 24GB GPU: What Actually Works in 2026.
Which model has the best Singaporean English accent retention?
In this benchmark, VoxCPM and IndexTTS2 both retained the FEMALE_01 accent profile well. Qwen3-TTS at the right scale also retained it. CosyVoice3 (current run) had inconsistent retention.
Are any of these models commercially licensed for production use?
License status varies and can change. Current official IndexTTS2 sources carry a custom Bilibili Model Use License with additional conditions. Qwen3-TTS and the checked VoxCPM code and weights identify Apache-2.0 licensing, while CosyVoice has its own upstream materials to review. Verify the current code, base-weight, adapted-weight, dataset, and voice-rights terms before deployment. This is not legal advice.
What's the difference between LoRA and full SFT for TTS finetuning?
LoRA (Low-Rank Adaptation) trains a small adapter while most base-model parameters stay frozen. Full SFT (Supervised Fine-Tuning) updates all or a much larger portion of the model. LoRA usually reduces trainable parameters and artifact size; full SFT increases adaptation capacity and memory risk. Neither is automatically better, so compare both against the same baseline when the backend supports them.