Datasets:
audio audioduration (s) 1.8 9.88 |
|---|
mytts English TTS: Tier-1 ablations
Archive: screening runs, not the released model. This repository keeps the ablation runs that chose the recipe of the main model: small models trained on a subset of the data, each technique compared with two baseline seeds. It is the record of those decisions; the checkpoints here are experiments.
- The model: VoiceHub/DACFlow-EN-10k (checkpoints, samples you can listen to, how to use it)
- Its training data (tokenized): VoiceHub/DACFlow-EN-10k-data
- Samples of these screening runs (archive too): VoiceHub/mytts-en-samples
Screening runs of the English roadmap in kadirnar/dacvae-next
(branch roadmap/en-echo, protocol EN-E1): every technique is trained with the small model (80 M, 30k steps, 12k frames per
step) on the Tier-1 subset en_t1 of SynDataLab-EN/echo-clones-4m-en
(strided shards of both generation runs, ~100 h, 236 voices held out), and scored on the two decision sets:
- echo_dev: 500 sentences voiced by held-out echo voices x prompt draws d0 + d1 (in-domain; SIM there is in-distribution);
- seed_dev: the frozen dev half of Seed-TTS test-en (Common Voice prompts, ~545 cases) x sampling seeds 0 + 1 (real voices).
Protocol: 32 Euler steps + sway, joint CFG w 4, initial noise 0.9, -16 LUFS, duration by the band rule (no oracle lengths);
ASR openai/whisper-large-v3 (greedy, English, repetition-loop guard) with the frozen en-v2 normalization (fillers,
apostrophes, spelling variants); corpus WER/CER; SIM-o = WavLM-large + ECAPA-TDNN cosine to the prompt (the Seed-TTS-eval /
F5-TTS scorer); UTMOS22 strong. Seed-TTS-eval and LibriSpeech-PC numbers comparable with papers are milestone-only
(scripts/recipes/en_final_eval.sh), never decision numbers.
Decision rule (EN-E1). A technique is adopted only when its paired corpus-WER gain excludes 0 against BOTH baseline
seeds (en-t1-base, en-t1-base-s2) on echo_dev or seed_dev, it is significantly worse than neither seed on the other set, SIM-o drops by at most
0.01 and UTMOS by no more than the seeds' own spread (the Turkish seeds differed by 1.5 WER points). Brackets are 95 % CIs of
the corpus-WER difference (points) from a paired bootstrap over sentences (echo) or prompt-connected components (Seed).
Results
| run | eval | step | WER | CER | word errors | SIM | UTMOS | ΔWER vs en-t1-base | ΔWER vs en-t1-base-s2 | changes vs baseline |
|---|---|---|---|---|---|---|---|---|---|---|
| en-t1-base | echo_dev | 30000 | 1.96% | 1.36% | 85/4344 | 0.631 | 2.68 | - | -0.55 [-1.36, +0.17] | (reference) |
| en-t1-base | seed_dev | 30000 | 5.96% | 2.73% | 713/11968 | 0.393 | 2.58 | - | +0.12 [-0.53, +0.77] | (reference) |
| en-t1-base-s2 | echo_dev | 30000 | 2.51% | 1.66% | 109/4344 | 0.630 | 2.70 | +0.55 [-0.17, +1.36] | - | train.seed=2 |
| en-t1-base-s2 | seed_dev | 30000 | 5.84% | 2.61% | 699/11968 | 0.392 | 2.62 | -0.12 [-0.77, +0.53] | - | train.seed=2 |
| en25-hubl18 | echo_dev | 30000 | 2.03% | 1.41% | 88/4344 | 0.607 | 2.59 | +0.07 [-0.55, +0.66] | -0.48 [-1.30, +0.20] | data.data_root=/workspace/data/tts_en_hubl18 |
| en25-hubl18 | seed_dev | 30000 | 5.49% | 2.59% | 657/11968 | 0.369 | 2.54 | -0.47 [-1.12, +0.15] | -0.35 [-0.98, +0.28] | data.data_root=/workspace/data/tts_en_hubl18 |
| en25-hubl24 | echo_dev | 30000 | 1.68% | 1.10% | 73/4344 | 0.617 | 2.59 | -0.28 [-0.96, +0.40] | -0.83 [-1.58, -0.07] | data.data_root=/workspace/data/tts_en_hubl24 |
| en25-hubl24 | seed_dev | 30000 | 5.37% | 2.26% | 643/11968 | 0.381 | 2.45 | -0.58 [-1.23, +0.01] | -0.47 [-1.13, +0.14] | data.data_root=/workspace/data/tts_en_hubl24 |
| en25-hubl24fp32 | echo_dev | 30000 | 1.75% | 1.12% | 76/4344 | 0.614 | 2.55 | -0.21 [-0.79, +0.37] | -0.76 [-1.49, -0.09] | data.data_root=/workspace/data/tts_en_hubl24_fp32 |
| en25-hubl24fp32 | seed_dev | 30000 | 5.82% | 2.60% | 696/11968 | 0.394 | 2.34 | -0.14 [-0.77, +0.46] | -0.03 [-0.73, +0.63] | data.data_root=/workspace/data/tts_en_hubl24_fp32 |
| en29-fixed | echo_dev | 30000 | 1.73% | 1.15% | 75/4344 | 0.639 | 2.76 | -0.23 [-0.90, +0.44] | -0.78 [-1.57, -0.07] | data.cross_plan=True, data.prompt_crop=True, data.prompt_picker=exact |
| en29-fixed | seed_dev | 30000 | 5.21% | 2.19% | 624/11968 | 0.392 | 2.65 | -0.74 [-1.35, -0.17] | -0.63 [-1.26, -0.02] | data.cross_plan=True, data.prompt_crop=True, data.prompt_picker=exact |
| en29-fixed-pc06 | echo_dev | 30000 | 1.06% | 0.68% | 46/4344 | 0.652 | 2.71 | -0.90 [-1.41, -0.42] | -1.45 [-2.18, -0.78] | data.cross_plan=True, data.p_cross=0.6, data.prompt_crop=True, data.prompt_picker=exact |
| en29-fixed-pc06 | seed_dev | 30000 | 6.30% | 2.85% | 754/11968 | 0.424 | 2.63 | +0.34 [-0.29, +0.95] | +0.46 [-0.20, +1.03] | data.cross_plan=True, data.p_cross=0.6, data.prompt_crop=True, data.prompt_picker=exact |
| en29-fixed-pc06-tf04 | echo_dev | 30000 | 1.24% | 0.74% | 54/4344 | 0.649 | 2.76 | -0.71 [-1.28, -0.18] | -1.27 [-1.98, -0.62] | data.cross_plan=True, data.p_cross=0.6, data.p_drop_prompt_text=0.4, data.prompt_crop=True, data.prompt_picker=exact |
| en29-fixed-pc06-tf04 | seed_dev | 30000 | 5.97% | 2.71% | 714/11968 | 0.407 | 2.58 | +0.01 [-0.61, +0.57] | +0.13 [-0.49, +0.70] | data.cross_plan=True, data.p_cross=0.6, data.p_drop_prompt_text=0.4, data.prompt_crop=True, data.prompt_picker=exact |
| en55-pc06-t06 | echo_dev | 30000 | 2.14% | 1.45% | 93/4344 | 0.687 | 3.09 | +0.18 [-0.50, +0.85] | -0.37 [-1.21, +0.40] | data.cross_plan=True, data.p_cross=0.6, data.prompt_crop=True, data.prompt_picker=exact, flow.t_mean=-0.6, flow.t_std=0.9 |
| en55-pc06-t06 | seed_dev | 30000 | 6.56% | 2.99% | 785/11968 | 0.449 | 2.99 | +0.60 [-0.04, +1.22] | +0.72 [+0.02, +1.38] | data.cross_plan=True, data.p_cross=0.6, data.prompt_crop=True, data.prompt_picker=exact, flow.t_mean=-0.6, flow.t_std=0.9 |
| en55-pc06-t08 | echo_dev | 30000 | 1.68% | 1.09% | 73/4344 | 0.681 | 3.04 | -0.28 [-1.01, +0.44] | -0.83 [-1.70, -0.05] | data.cross_plan=True, data.p_cross=0.6, data.prompt_crop=True, data.prompt_picker=exact, flow.t_mean=-0.8 |
| en55-pc06-t08 | seed_dev | 30000 | 5.30% | 2.31% | 634/11968 | 0.440 | 2.94 | -0.66 [-1.40, -0.01] | -0.54 [-1.23, +0.14] | data.cross_plan=True, data.p_cross=0.6, data.prompt_crop=True, data.prompt_picker=exact, flow.t_mean=-0.8 |
| en55-pc06-t10 | echo_dev | 30000 | 2.07% | 1.45% | 90/4344 | 0.663 | 2.95 | +0.12 [-0.53, +0.75] | -0.44 [-1.17, +0.27] | data.cross_plan=True, data.p_cross=0.6, data.prompt_crop=True, data.prompt_picker=exact, flow.t_mean=-1.0 |
| en55-pc06-t10 | seed_dev | 30000 | 5.21% | 2.30% | 624/11968 | 0.413 | 2.79 | -0.74 [-1.40, -0.15] | -0.63 [-1.28, +0.01] | data.cross_plan=True, data.p_cross=0.6, data.prompt_crop=True, data.prompt_picker=exact, flow.t_mean=-1.0 |
| en55-repafade | echo_dev | 30000 | 3.15% | 2.06% | 137/4344 | 0.628 | 2.70 | +1.20 [+0.48, +1.92] | +0.64 [-0.18, +1.44] | train.repa_fade_steps=3000, train.repa_until_step=18000 |
| en55-repafade | seed_dev | 30000 | 7.28% | 3.53% | 871/11968 | 0.398 | 2.51 | +1.32 [+0.60, +2.00] | +1.44 [+0.73, +2.11] | train.repa_fade_steps=3000, train.repa_until_step=18000 |
| en55-t06 | echo_dev | 30000 | 3.34% | 2.27% | 145/4344 | 0.676 | 3.09 | +1.38 [+0.73, +2.03] | +0.83 [-0.09, +1.60] | flow.t_mean=-0.6, flow.t_std=0.9 |
| en55-t06 | seed_dev | 30000 | 8.01% | 3.87% | 959/11968 | 0.448 | 2.96 | +2.06 [+1.36, +2.71] | +2.17 [+1.40, +2.92] | flow.t_mean=-0.6, flow.t_std=0.9 |
| en55-t08 | echo_dev | 30000 | 3.38% | 2.43% | 147/4344 | 0.662 | 2.93 | +1.43 [+0.51, +2.35] | +0.87 [+0.00, +1.79] | flow.t_mean=-0.8 |
| en55-t08 | seed_dev | 30000 | 8.01% | 3.79% | 959/11968 | 0.435 | 2.78 | +2.06 [+1.34, +2.74] | +2.17 [+1.42, +2.89] | flow.t_mean=-0.8 |
Decisions (EN-E1)
| run | verdict | WER gain vs both seeds on | worse than a seed on | guard failures |
|---|---|---|---|---|
| en-t1-base | no significant gain (provisional: baselines en-t1-base-s2) | - | - | - |
| en-t1-base-s2 | no significant gain (provisional: baselines en-t1-base) | - | - | - |
| en25-hubl18 | no significant gain | - | - | SIM-o on echo_dev (-0.024, -0.022); SIM-o on seed_dev (-0.024, -0.023); UTMOS on echo_dev (2.59 vs seed mean 2.69, spread 0.02); UTMOS on seed_dev (2.54 vs seed mean 2.60, spread 0.04) |
| en25-hubl24 | no significant gain | - | - | SIM-o on echo_dev (-0.014, -0.013); SIM-o on seed_dev (-0.011, -0.010); UTMOS on echo_dev (2.59 vs seed mean 2.69, spread 0.02); UTMOS on seed_dev (2.45 vs seed mean 2.60, spread 0.04) |
| en25-hubl24fp32 | no significant gain | - | - | SIM-o on echo_dev (-0.017, -0.016); UTMOS on echo_dev (2.55 vs seed mean 2.69, spread 0.02); UTMOS on seed_dev (2.34 vs seed mean 2.60, spread 0.04) |
| en29-fixed | adopt | seed_dev | - | - |
| en29-fixed-pc06 | adopt | echo_dev | - | - |
| en29-fixed-pc06-tf04 | adopt | echo_dev | - | - |
| en55-pc06-t06 | worse on seed_dev | - | seed_dev | - |
| en55-pc06-t08 | no significant gain | - | - | - |
| en55-pc06-t10 | no significant gain | - | - | - |
| en55-repafade | worse on echo_dev, seed_dev | - | echo_dev, seed_dev | UTMOS on seed_dev (2.51 vs seed mean 2.60, spread 0.04) |
| en55-t06 | worse on echo_dev, seed_dev | - | echo_dev, seed_dev | - |
| en55-t08 | worse on echo_dev, seed_dev | - | echo_dev, seed_dev | - |
- Downloads last month
- 1,675