TL;DR
Published 2026-08-24. Fine-tuned Dots.TTS soar ranked first, ahead of fine-tuned MOSS-TTS v1.5 and OmniVoice + LoRA checkpoint 800. Among zero-shot arms, Dots.TTS soar ranked highest (5) and MOSS-TTS v1.5 zero-shot ranked 13.
- Published
- 2026-08-24
- Published model rows
- 16
- Evidence
- Historical task evaluation
Follow this evaluation
Published results
Voice-cloning systems ranked blind on one speaker's voice: source evidence dated 2026-08-24.
- Exact model version
- Dots.TTS soar, fine-tuned
- Blind rank
- 1
- Licence
- Apache 2.0, 2.2B
- Verdict (author, blind)
- "strong", "great, no crazy defects"
- Reported clip measurement
- 1.3 seconds take-to-take spread across five renders
The displayed values belong to the identified model versions and this published task, with their original metric units.
What did this task require the model to return?
Read the tested version, task and rubric together. Check failures alongside successful answers; compare costs only within the reported unit.
2026-08-24: Voice-cloning systems ranked blind on one speaker's voiceKeep the study date and exact model ID beside the result. Do not average different tasks into one winner.
Test a strategy with Aurora
Keep the study date and exact model ID beside the result. Do not average different tasks into one winner.
Test a strategy with AuroraKeep reporting dates and units beside the answer, and inspect its original citations.
Published results
See the complete published results
Scroll sideways to compare all columns. Select a heading to sort.
| Dots.TTS soar, fine-tuned | 1 | Apache 2.0, 2.2B | "strong", "great, no crazy defects" | 1.3 seconds take-to-take spread across five renders |
| MOSS-TTS v1.5, fine-tuned | 2 | Apache 2.0, 8.5B | "great", "extremely good... it doesn't collapse" | — |
| OmniVoice + LoRA, checkpoint 800 | 3 | open code, CC-BY-NC weights | "probably the best... but end of sentences still collapse hard" | — |
| CosyVoice 3 0.5B, fine-tuned | 4 | Apache 2.0 | "really good, no issues" | — |
| Dots.TTS soar, zero-shot | 5 | Apache 2.0 | "pretty good" | 98.6% word coverage, zero added words, 23.84s |
| OmniVoice, zero-shot | 6 | open code, CC-BY-NC weights | "really good", "decent" | 100% word coverage after trimming 2.0s of hallucinated speech |
| Audio8 0.6b, zero-shot | 7 | Apache 2.0 | "okay", "good-ish" | 100% word coverage, 27.31s |
| Chatterbox, zero-shot | 8 | MIT | "ok" | — |
| Qwen3-TTS, zero-shot | 9 | Apache 2.0 | "okay" | — |
| IndexTTS-2, zero-shot | 10 | custom model licence | "okay, robotic, too fast" | — |
| CosyVoice 3, zero-shot | 11 | Apache 2.0 | "okay" | — |
| Audio8 0.1b, zero-shot | 12 | Audio8 Community License v1.0 | "okay, robotic" | 100% word coverage, 27.26s |
| MOSS-TTS v1.5, zero-shot (8.5B) | 13 | Apache 2.0 | "decent, worse than Audio8 0.6b" | 21.76s |
| ElevenLabs Professional Voice Clone | 14 | $22/mo | close, not close enough | — |
| Qwen3-TTS 1.7B, fine-tuned | 15 | Apache 2.0 | great audio, not my voice | — |
| XTTS-v2, fine-tuned | 16 | non-commercial | "hilariously bad" | — |
No matching rows. Clear the filter to see all records.
16 ranked arms (fine-tuned and zero-shot variants of what the article calls eleven systems), each reading the same 67-word script, plus a recorded human control. Published 2026-08-24. Values retain their original units and grading scale.
Task and tested input
Clone one speaker's voice and read a fixed script so it sounds like that speaker.
Instruction in brief (task summary)
Every arm reads the identical 67-word script. Fine-tuned arms train on a 45-minute single-session phone recording. Zero-shot arms condition on the same clean 20-second reference.
How the runs worked
16 ranked arms (fine-tuned and zero-shot variants of what the article calls eleven systems), each reading the same 67-word script, plus a recorded human control
Author-run blind listening sessions with loudness-matched, duration-padded clips and a sealed label mapping
Step 1
Record about 45 minutes of speech in one session and build a 24 kHz clip corpus.
Step 2
Fine-tune or zero-shot render each system on the same 67-word script.
Step 3
Normalise loudness, pad to one duration and seal the label mapping.
Step 4
Rank blind across sessions, then separate the top two over three sealed rounds including an unseen script with three takes per model.
How to read the measurements
Grading and valid output
The author's blind subjective ranking of similarity to his own voice. Word coverage figures come from Whisper, which the article says it no longer uses to choose takes.
Cost
The article publishes no per-model inference cost. It reports ElevenLabs Creator at $22/mo, fine-tuning at under $1 per run, and about $17 of GPU time for the whole project.
Time
The article publishes no per-model timing. Durations in the table are clip lengths, not generation time.
Why this model was selected
The article selected fine-tuned Dots.TTS soar as the model it ships, citing its Apache 2.0 licence, 2.2B size and wins in the final rounds. It named OmniVoice as closest to the author's voice but unshippable because of CC-BY-NC weights and collapsing phrase endings.
See the published results
This image was published with the 2026-08-24 article. Read the source for the full test procedure and every result it reports.
Sample and scope
The table merges blind rounds run over four days. Only the top three met directly, and the article says to read adjacent lower rows as ties.
One listener (the author) on one voice, with one published take per arm. The article shows the same checkpoint varying from unusable to good across takes.
August 2026 versions with unequal inputs: fine-tunes had 45 minutes of audio, zero-shot arms had 20 seconds, and ElevenLabs used cleaner source audio.
No comments yet.