DACVAE-TTS Turkish: every tr-combined experiment in one place

Training on Codyfederer/tr-combined, code and write-up: https://github.com/kadirnar/dacvae-tts (branch trc/tr-combined-experiments). 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).

Current result: new models vs run C (single sample, no reranking)

Same inference for all three: duration predictor refit on tr-combined, guidance 5, 32 steps, 495 Freya-TR-Eval sentences × 48 leak-free Common Voice voices, sampling seeds 42 + 1000 pooled. Brackets: paired speaker-clustered jackknife-t 95 % interval of the difference to run C.

model WER % CER % SIM-o DNSMOS UTMOS
run C (VoiceHub/dacvae-tts-tr-w512, old data, 70 h) 5.10 2.93 0.519 2.860 2.493
full-cross: tr-combined + cross-utterance prompts, 60k 2.94 [-3.27, -1.05] 1.66 [-2.18, -0.34] 0.536 2.936 2.572
full-v2: + quality condition + speech-REPA + character units, 60k 0.93 [-5.35, -2.99] 0.36 [-3.52, -1.62] 0.556 3.128 2.533

full-v2 against full-cross: WER -2.01 [-2.55, -1.46], CER -1.31 [-1.66, -0.95], SIM-o +0.021 [0.010, 0.031], DNSMOS +0.192 [0.163, 0.221], UTMOS -0.039 [-0.084, 0.006] (tie). Seed 42: 468 of 495 sentences have no word error; the rest are near-homophones ("Vurmak da" → "Vurmakta"), foreign names and one repeated phrase.

What was adopted, from the 20k A/B arms below: cross-utterance prompts (#11), speech-REPA (#10: WER halved, alignment ~4x earlier), the quality condition (DNSMOS control), character units (small, consistent over two training seeds). Rejected: latent negatives (#8, collapse), TLA-SA (#10), model-guidance fine-tune (#14), Flow-GRPO at 600 updates (#16, no effect). Side-by-side audio: comparison. Inference comparisons: systems.

Final evaluation of every run

Freya-TR-Eval (495 sentences) spoken by 48 leak-free Common Voice test voices, one sample per sentence (no reranking), guidance 5, 32 Euler steps, prompt-rate duration rule, Whisper large-v3, turkish-v2 metric. Seeds: sampling seeds of the final evaluation (2 = 42 and 1000 pooled). A/B arms stop at 20k of the 60k schedule; training-seed spread there is ~5 WER points.

run what updates seeds WER % CER % SIM-o DNSMOS UTMOS
full-cross model candidate: run C recipe + cross prompts, 60k 60000 2 6.46 4.40 0.548 2.970 2.519
full-v2 model candidate v2: cross + quality + REPA + char units, 60k: quality_condition=true, quality_scores=quality/dnsmos.json, quality_target=[4.0, 4.5, 3.8], re... 60000 2 2.21 1.17 0.577 3.164 2.515
run-c-reference run C (VoiceHub/dacvae-tts-tr-w512, old data Vyvo/tr-dataset-12), same protocol 60000 1 7.54 4.89 0.535 2.903 2.420
y-base 20000 2 4.32 2.98 0.587 3.111 2.486
y-cond-text-pool #9, on v2: cond_text_pool=true 20000 2 4.53 3.07 0.590 3.112 2.499
y-decay-matrices audit: weight decay on matrices only, on v2: weight_decay_scope=matrices 20000 2 3.77 2.11 0.581 3.122 2.480
y-ffn-conv #9, on v2: ffn_conv_kernel=5 20000 2 4.04 2.74 0.598 3.090 2.415
y-final-adaln #9, on v2: final_adaln=true 20000 2 4.04 2.27 0.588 3.128 2.495
y-long-skip #9, on v2: long_skip=true 20000 2 3.99 2.76 0.589 3.110 2.512
y-regularized #14 dropout 0.1 + weight decay 0.05 on matrices, on v2: dropout=0.1, weight_decay=0.05, weight_decay_scope=matrices 20000 2 4.58 3.05 0.572 3.043 2.397
y-s43 noise floor of the v2 recipe: seed=43 20000 2 3.64 2.30 0.571 3.112 2.506
y-speaker-condition round 2, on v2: speaker_condition_dim=192, speaker_condition=teacher/ecapa-speechbrain 20000 2 3.66 2.40 0.599 3.039 2.295
y-value-residual #9, on v2: value_residual=true 20000 2 3.94 2.29 0.582 3.138 2.533
ft-mg-w07 #14 model guidance w=0.7 (sampled at g=1) 8000 2 9.01 5.82 0.526 3.022 2.546
ft-w0 #14 control: same fine-tune, w=0 (sampled at g=5) 8000 2 6.58 4.42 0.546 2.964 2.486
grpo #16 Flow-GRPO from full-cross 60k 600 2 6.70 4.51 0.551 2.976 2.536
base-eager #7 20000 2 20.12 13.12 0.553 2.972 2.430
base-s42 #7 baseline 20000 2 30.09 22.11 0.516 2.925 2.312
base-s42-pad64 #7 baseline with lengths padded to multiples of 64 (20 % of every batch was padding) 20000 2 27.87 17.76 0.534 2.886 2.303
base-s43 seed noise floor: seed=43 20000 2 24.90 15.02 0.543 2.884 2.329
pairs-cross #11: cross_prompt_prob=0.4, cross_prompt_max_utterances=3, cross_prompt_max_seconds=12.0 20000 2 12.30 8.25 0.540 2.961 2.445
x-attn-gate #9, on base+cross: attn_gate=head 20000 2 16.67 10.25 0.529 2.917 2.375
x-char-units round 2, on base+cross: text_units=chars 20000 2 11.17 7.41 0.551 2.973 2.439
x-char-units-s43 round 2 char units, training seed 43: text_units=chars, seed=43 20000 2 16.08 9.71 0.546 2.926 2.341
x-latent-negatives #8, on base+cross: contrastive_mode=latent_delta, contrastive_random_weight=0.2, contrastive_aug_weight=0.2, contrastive_span_min=3, contrastive_span_max=12... 20000 2 62.15 38.60 0.139 1.919 1.258
x-long-skip #9, on base+cross: long_skip=true 20000 2 13.01 8.21 0.535 2.953 2.407
x-no-negatives #8, on base+cross: contrastive_mode=none 20000 2 13.92 9.90 0.540 2.965 2.430
x-pairs-char-ctc #11, on base+cross: ctc_targets=chars 20000 2 13.64 8.73 0.540 2.939 2.429
x-pairs-tail #11, on base+cross: tail_silence_prob=0.3, tail_silence_max_seconds=0.8, long_prompt_prob=0.25, prompt_fraction_long_max=0.85, prompt_cut=quiet 20000 2 13.26 8.89 0.543 2.972 2.438
x-quality-cond new: quality condition, on base+cross: quality_condition=true, quality_scores=quality/dnsmos.json 20000 2 13.22 8.67 0.533 3.020 2.491
x-repa #10, on base+cross: repa_layer=10, repa_dim=256, teacher_features=teacher/mhubert147-l12-pca256, repa_weight=1.0, repa_stop_step=0, repa_frames=all 20000 2 5.56 3.76 0.585 2.882 2.207
x-s43 noise floor of base+cross: seed=43 20000 2 16.94 9.66 0.533 2.936 2.394
x-swiglu #9, on base+cross: ffn_activation=swiglu 20000 2 17.45 9.91 0.524 2.903 2.366
x-tla #10, on base+cross: tla_layers=all, tla_dim=192, tla_hidden=256, speaker_embeddings=teacher/ecapa-speechbrain, tla_weight=0.5, tla_entropy=0.01 20000 2 24.67 14.05 0.455 2.771 2.167

Training trajectories

WER / CER % per snapshot: the first 96 sentences (quick check) before the final step, seed 42.

run 5k 10k 15k 20k 25k 30k 35k 40k 45k 50k 55k 60k
full-cross 92.9 / 53.8 27.9 / 16.5 17.9 / 11.5 12.0 / 8.5 11.4 / 7.2 9.9 / 6.5 9.2 / 6.2 7.2 / 4.8 8.1 / 4.9 8.3 / 5.9 7.9 / 6.2 6.7 / 4.7
full-v2 11.3 / 6.7 6.8 / 4.9 5.8 / 4.5 4.3 / 3.7 5.1 / 3.3 3.4 / 2.4 2.4 / 1.6 2.6 / 1.8 2.3 / 1.7 2.2 / 1.3 1.8 / 1.1 2.4 / 1.4
y-base 11.3 / 6.7 6.8 / 4.9 5.8 / 4.5 4.3 / 2.9
y-cond-text-pool 12.2 / 8.5 5.5 / 3.3 5.5 / 3.6 4.4 / 2.9
y-decay-matrices 12.3 / 7.7 7.4 / 5.1 5.0 / 3.0 3.7 / 2.0
y-ffn-conv 10.8 / 6.5 6.0 / 4.4 4.9 / 3.4 4.2 / 2.8
y-final-adaln 12.4 / 9.2 7.5 / 5.2 5.5 / 3.4 4.2 / 2.3
y-long-skip 12.3 / 8.9 7.3 / 5.2 5.5 / 3.4 3.9 / 2.9
y-regularized 10.7 / 6.7 4.4 / 3.5 6.4 / 4.7 4.8 / 3.2
y-s43 11.4 / 6.3 7.5 / 4.7 4.8 / 3.5 3.6 / 2.3
y-speaker-condition 10.1 / 6.2 7.2 / 5.1 4.1 / 3.5 3.7 / 2.4
y-swiglu 12.6 / 7.7 7.1 / 5.1 6.5 / 4.4
y-value-residual 10.8 / 6.8 6.4 / 4.1 7.3 / 4.4 3.6 / 2.0
base-eager 73.3 / 51.6 40.0 / 25.9 28.7 / 19.7 19.8 / 12.7
base-s42 76.4 / 46.8 36.4 / 21.4 32.2 / 24.5 30.3 / 22.1
base-s42-pad64 83.6 / 48.8 52.7 / 32.7 40.5 / 25.2 27.6 / 17.1
base-s43 46.0 / 27.3 31.5 / 18.4 24.5 / 15.0
pairs-cross 97.6 / 58.5 29.6 / 17.6 17.3 / 11.9 12.0 / 8.2
x-attn-gate 99.1 / 56.3 29.3 / 16.9 18.8 / 12.0 16.8 / 10.3
x-char-units 93.8 / 55.8 30.6 / 18.1 16.3 / 9.9 11.1 / 7.4
x-char-units-s43 85.6 / 48.7 33.6 / 19.2 21.0 / 13.3 16.2 / 9.8
x-latent-negatives 104.2 / 86.4 95.0 / 64.7 75.1 / 48.2 61.9 / 38.8
x-long-skip 90.5 / 56.4 30.6 / 18.3 17.4 / 10.9 12.9 / 8.1
x-no-negatives 101.9 / 59.3 34.5 / 19.5 17.9 / 12.3 14.4 / 10.1
x-pairs-char-ctc 98.7 / 66.6 26.7 / 16.4 19.4 / 12.3 13.6 / 8.6
x-pairs-tail 95.4 / 59.3 28.0 / 16.4 17.5 / 11.8 13.0 / 8.9
x-quality-cond 97.8 / 67.9 29.3 / 17.3 18.5 / 11.6 13.6 / 8.9
x-repa 13.6 / 8.9 10.5 / 7.9 7.2 / 5.0 6.1 / 3.9
x-s43 101.7 / 58.1 36.7 / 19.0 29.3 / 16.1 17.2 / 9.6
x-swiglu 98.6 / 58.1 38.7 / 22.2 24.7 / 19.3 17.8 / 10.0
x-tla 112.3 / 83.0 43.5 / 25.2 29.6 / 17.7 24.2 / 13.4

Other folders

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support