/ Модель
Каталог

Характеристики

Опубликована
26 сен 2026

Ссылки

Hugging Face

dacvae-tts-tr-combined

VoiceHub
Синтез речиТекст → аудио

Экспериментальная модель VoiceHub преобразует турецкий текст в речь и исследует синтез с клонированием голоса.

0скачиваний за 30 дней1лайквеса обновлены26 сен 2026

Бенчмарки

Как модель справилась с задачами пользователей ailist и что о ней публикуют открытые лидерборды

Публичные оценки

Из открытых лидербордов. Сравнивайте результаты одного теста при одинаковых условиях — отсутствие оценки не значит, что модель хуже.

В подключённых источниках результатов для этой модели пока нет.

Что умеет модель

Описание составлено по карточке модели и метаданным Hugging Face

DACVAE-TTS Turkish — экспериментальная модель синтеза речи на турецком языке, разработанная VoiceHub. Репозиторий объединяет результаты экспериментов с обучением на Codyfederer/tr-combined: для запусков представлены контрольные точки, конфигурации, журналы и оценки. В README описана проверка на 495 предложениях и 48 голосах Common Voice, с оценкой распознанных Whisper large-v3 транскриптов по WER и CER.

Модель предназначена для преобразования турецкого текста в речь; теги также указывают на поддержку клонирования голоса и применение flow matching. В оценке использовались guidance 5 и 32 шага Euler. Сведения о размере модели, длине контекста и лицензии в предоставленных данных отсутствуют. Результаты приведены для отдельных экспериментальных запусков и не гарантируют качество на других текстах и голосах.

Сценарии применения

  • синтез турецкой речи из текста
  • создание озвучки турецких фраз
  • эксперименты с клонированием турецкого голоса

Ограничения

README описывает экспериментальные оценки на ограниченном наборе предложений и голосов; качество вне этих условий не установлено. Лицензия и размер модели в предоставленных данных не указаны.

Карточка модели

README от автора на Hugging Face, как он его опубликовал

Открыть на Hugging Face

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.

modelWER %CER %SIM-oDNSMOSUTMOS
run C (VoiceHub/dacvae-tts-tr-w512, old data, 70 h)5.102.930.5192.8602.493
full-cross: tr-combined + cross-utterance prompts, 60k2.94 [-3.27, -1.05]1.66 [-2.18, -0.34]0.5362.9362.572
full-v2: + quality condition + speech-REPA + character units, 60k0.93 [-5.35, -2.99]0.36 [-3.52, -1.62]0.5563.1282.533

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.

runwhatupdatesseedsWER %CER %SIM-oDNSMOSUTMOS
full-crossmodel candidate: run C recipe + cross prompts, 60k6000026.464.400.5482.9702.519
full-v2model 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...6000022.211.170.5773.1642.515
run-c-referencerun C (VoiceHub/dacvae-tts-tr-w512, old data Vyvo/tr-dataset-12), same protocol6000017.544.890.5352.9032.420
y-base2000024.322.980.5873.1112.486
y-cond-text-pool#9, on v2: cond_text_pool=true2000024.533.070.5903.1122.499
y-decay-matricesaudit: weight decay on matrices only, on v2: weight_decay_scope=matrices2000023.772.110.5813.1222.480
y-ffn-conv#9, on v2: ffn_conv_kernel=52000024.042.740.5983.0902.415
y-final-adaln#9, on v2: final_adaln=true2000024.042.270.5883.1282.495
y-long-skip#9, on v2: long_skip=true2000023.992.760.5893.1102.512
y-regularized#14 dropout 0.1 + weight decay 0.05 on matrices, on v2: dropout=0.1, weight_decay=0.05, weight_decay_scope=matrices2000024.583.050.5723.0432.397
y-s43noise floor of the v2 recipe: seed=432000023.642.300.5713.1122.506
y-speaker-conditionround 2, on v2: speaker_condition_dim=192, speaker_condition=teacher/ecapa-speechbrain2000023.662.400.5993.0392.295
y-swiglu#9, on v2: ffn_activation=swiglu2000024.132.780.5693.0922.509
y-value-residual#9, on v2: value_residual=true2000023.942.290.5823.1382.533
ft-mg-w07#14 model guidance w=0.7 (sampled at g=1)800029.015.820.5263.0222.546
ft-w0#14 control: same fine-tune, w=0 (sampled at g=5)800026.584.420.5462.9642.486
grpo#16 Flow-GRPO from full-cross 60k60026.704.510.5512.9762.536
base-eager#720000220.1213.120.5532.9722.430
base-s42#7 baseline20000230.0922.110.5162.9252.312
base-s42-pad64#7 baseline with lengths padded to multiples of 64 (20 % of every batch was padding)20000227.8717.760.5342.8862.303
base-s43seed noise floor: seed=4320000224.9015.020.5432.8842.329
pairs-cross#11: cross_prompt_prob=0.4, cross_prompt_max_utterances=3, cross_prompt_max_seconds=12.020000212.308.250.5402.9612.445
x-attn-gate#9, on base+cross: attn_gate=head20000216.6710.250.5292.9172.375
x-char-unitsround 2, on base+cross: text_units=chars20000211.177.410.5512.9732.439
x-char-units-s43round 2 char units, training seed 43: text_units=chars, seed=4320000216.089.710.5462.9262.341
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...20000262.1538.600.1391.9191.258
x-long-skip#9, on base+cross: long_skip=true20000213.018.210.5352.9532.407
x-no-negatives#8, on base+cross: contrastive_mode=none20000213.929.900.5402.9652.430
x-pairs-char-ctc#11, on base+cross: ctc_targets=chars20000213.648.730.5402.9392.429
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=quiet20000213.268.890.5432.9722.438
x-quality-condnew: quality condition, on base+cross: quality_condition=true, quality_scores=quality/dnsmos.json20000213.228.670.5333.0202.491
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=all2000025.563.760.5852.8822.207
x-s43noise floor of base+cross: seed=4320000216.949.660.5332.9362.394
x-swiglu#9, on base+cross: ffn_activation=swiglu20000217.459.910.5242.9032.366
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=0.0120000224.6714.050.4552.7712.167

Training trajectories

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

run5k10k15k20k25k30k35k40k45k50k55k60k
full-cross92.9 / 53.827.9 / 16.517.9 / 11.512.0 / 8.511.4 / 7.29.9 / 6.59.2 / 6.27.2 / 4.88.1 / 4.98.3 / 5.97.9 / 6.26.7 / 4.7
full-v211.3 / 6.76.8 / 4.95.8 / 4.54.3 / 3.75.1 / 3.33.4 / 2.42.4 / 1.62.6 / 1.82.3 / 1.72.2 / 1.31.8 / 1.12.4 / 1.4
y-base11.3 / 6.76.8 / 4.95.8 / 4.54.3 / 2.9
y-cond-text-pool12.2 / 8.55.5 / 3.35.5 / 3.64.4 / 2.9
y-decay-matrices12.3 / 7.77.4 / 5.15.0 / 3.03.7 / 2.0
y-ffn-conv10.8 / 6.56.0 / 4.44.9 / 3.44.2 / 2.8
y-final-adaln12.4 / 9.27.5 / 5.25.5 / 3.44.2 / 2.3
y-long-skip12.3 / 8.97.3 / 5.25.5 / 3.43.9 / 2.9
y-regularized10.7 / 6.74.4 / 3.56.4 / 4.74.8 / 3.2
y-s4311.4 / 6.37.5 / 4.74.8 / 3.53.6 / 2.3
y-speaker-condition10.1 / 6.27.2 / 5.14.1 / 3.53.7 / 2.4
y-swiglu12.6 / 7.77.1 / 5.16.5 / 4.43.9 / 2.6
y-value-residual10.8 / 6.86.4 / 4.17.3 / 4.43.6 / 2.0
base-eager73.3 / 51.640.0 / 25.928.7 / 19.719.8 / 12.7
base-s4276.4 / 46.836.4 / 21.432.2 / 24.530.3 / 22.1
base-s42-pad6483.6 / 48.852.7 / 32.740.5 / 25.227.6 / 17.1
base-s4346.0 / 27.331.5 / 18.424.5 / 15.0
pairs-cross97.6 / 58.529.6 / 17.617.3 / 11.912.0 / 8.2
x-attn-gate99.1 / 56.329.3 / 16.918.8 / 12.016.8 / 10.3
x-char-units93.8 / 55.830.6 / 18.116.3 / 9.911.1 / 7.4
x-char-units-s4385.6 / 48.733.6 / 19.221.0 / 13.316.2 / 9.8
x-latent-negatives104.2 / 86.495.0 / 64.775.1 / 48.261.9 / 38.8
x-long-skip90.5 / 56.430.6 / 18.317.4 / 10.912.9 / 8.1
x-no-negatives101.9 / 59.334.5 / 19.517.9 / 12.314.4 / 10.1
x-pairs-char-ctc98.7 / 66.626.7 / 16.419.4 / 12.313.6 / 8.6
x-pairs-tail95.4 / 59.328.0 / 16.417.5 / 11.813.0 / 8.9
x-quality-cond97.8 / 67.929.3 / 17.318.5 / 11.613.6 / 8.9
x-repa13.6 / 8.910.5 / 7.97.2 / 5.06.1 / 3.9
x-s43101.7 / 58.136.7 / 19.029.3 / 16.117.2 / 9.6
x-swiglu98.6 / 58.138.7 / 22.224.7 / 19.317.8 / 10.0
x-tla112.3 / 83.043.5 / 25.229.6 / 17.724.2 / 13.4

Other folders

Как запустить

На своей GPU в ailist. Цена — по текущим предложениям провайдеров

GPU в ailist

Универсальный рантайм

Нельзя запустить

В репозитории нет config.json: не понять архитектуру модели.

Ревизия весов
b0a0dbf

0 моделей

    Создать бенчмарк