DACVAE-TTS English: every A/B arm in one place
Training on Vyvo/en-podcast-1, code and write-up: https://github.com/kadirnar/dacvae-tts. One folder per run: checkpoints/step-*.pt, eval/step-*/ (scores, per-sentence Whisper transcripts, audio), train.jsonl, config.json, its own README, RESULT.md (question, result against its baseline, verdict) and logs/.
EXPERIMENTS.md: the result log of every experiment (question, evidence, paired result against its baseline, trajectory, verdict and why).
Best model
full-base-s42 at 60,000 updates (checkpoint): 67.4M parameters,
configs/en_base.yaml, trained only on en-podcast-1 (120 h after filtering). Try it in the
demo.
| Seed-TTS test-en DEV half (545 sentences, seeds 42 + 1000) | WER % | CER % | SIM-o | UTMOS22 | DNSMOS |
|---|---|---|---|---|---|
full-base-s42, 60k | 2.63 | 1.29 | 0.599 | 2.995 | 2.535 |
| DACVAE round trip of the ground truth (ceiling) | 2.14 | 0.84 | 0.721 | 3.481 | 3.010 |
Listen: zero-shot voice cloning
20 clean reference voices that no training data contains (14 male, the rest female; EchoTTS speakers from SynDataLab-EN-Refs/echo-ref-speakers-4k-en, 44.1 kHz, chosen for the highest DNSMOS), each with its transcript, and a different sentence in each voice, synthesized by the snapshot at its operating point (32 Euler steps, guidance 5, prompt-rate duration). Single samples, seed 42, no reranking.
full-base-s42 at 60,000 updates
| # | voice | sentence | reference voice | generated |
|---|---|---|---|---|
| 1 | male, 110 Hz, ref DNSMOS 3.63 | Are they twins or just normal siblings? | ||
| 2 | male, 103 Hz, ref DNSMOS 3.55 | He must be disguised to avoid encounters with thieves. | ||
| 3 | female, 187 Hz, ref DNSMOS 3.53 | The town of Norwich is located east of the center of the county. | ||
| 4 | male, 142 Hz, ref DNSMOS 3.61 | Constrictor knots can also be quite effective as improvised hose clamps or cable ties. | ||
| 5 | male, 128 Hz, ref DNSMOS 3.55 | These two agencies co-operate in aspects of law enforcement, including their joint marine unit. | ||
| 6 | female, 173 Hz, ref DNSMOS 3.63 | It then became part of Poland. | ||
| 7 | male, 102 Hz, ref DNSMOS 3.60 | Will you be happy when I am in the earth? | ||
| 8 | male, 121 Hz, ref DNSMOS 3.60 | She was involved in several record productions. | ||
| 9 | female, 186 Hz, ref DNSMOS 3.61 | The Larson Brothers Airport is located in the town. | ||
| 10 | male, 116 Hz, ref DNSMOS 3.60 | Flood gates were erected and drainage systems improved. | ||
| 11 | male, 119 Hz, ref DNSMOS 3.59 | He did not live to complete the development of his system. | ||
| 12 | female, 190 Hz, ref DNSMOS 3.54 | The land is volcanic in origin with mountains in the interior. | ||
| 13 | male, 141 Hz, ref DNSMOS 3.53 | The Tuesday market is held by the blue bridge over the River Iro. | ||
| 14 | male, 96 Hz, ref DNSMOS 3.56 | The evidence against Strafford was weak, and the impeachment failed. | ||
| 15 | female, 188 Hz, ref DNSMOS 3.57 | It is operated by the New York City Department of Parks and Recreation. | ||
| 16 | male, 128 Hz, ref DNSMOS 3.57 | Also that year, he collaborated with Evan Dando on a new Lemonheads album. | ||
| 17 | male, 129 Hz, ref DNSMOS 3.53 | Finally, lumping groups of one's own pieces together is usually poor strategy. | ||
| 18 | female, 228 Hz, ref DNSMOS 3.59 | Over time the couple has four more children, although William never warms to them. | ||
| 19 | male, 139 Hz, ref DNSMOS 3.54 | The battle was of medium term significance, consolidating Royalist control of Yorkshire. | ||
| 20 | male, 99 Hz, ref DNSMOS 3.56 | Before guns were invented, armies had to throw bullets at each other and if a bullet touched you, you had to sit out until the next war. |
Everything else
Every run, checkpoint and evaluation is in the repository files, one folder per run (base-s42, base-s43, charts, continue60k-lowlr-s42, en-baseline, en-corpus-build, en-duration, full-base-s42, liam-ft-base60k-s42, liam-posttrain-20260928, repa-hubert-large, repa-layer4, repa-layer6, repa-layer6-s43, repa-off, repa-off-s43, repa-target, repa-w05, repa-w2), and every experiment's result is in EXPERIMENTS.md.