VoiceHub/dacvae-tts-tr-w512-clean
DACVAE-TTS Turkish run C (width 512, clean data) Generated audio of every evaluated checkpoint of the training run tr-w512-clean (Turkish zero-shot voice-cloning TTS, dacvae-tts, frozen Meta DACVAE latents, 48 kHz). This repository holds model outputs and metrics, not training data. Training data: Vyvo/tr-dataset-12 (Turkish podcast segments). Run C: configs/nano_tr_w512.yaml (66.5M parameters: width 512, 8 heads, batch expansion 4, frame budget 6000), trained from scratch on… See the full description on the dataset page: https://huggingface.co/datasets/VoiceHub/dacvae-tts-tr-w512-clean.
DACVAE-TTS Turkish run C (width 512, clean data)
Generated audio of every evaluated checkpoint of the training run tr-w512-clean (Turkish zero-shot voice-cloning TTS, dacvae-tts, frozen Meta DACVAE latents, 48 kHz). This repository holds model outputs and metrics, not training data. Training data: Vyvo/tr-dataset-12 (Turkish podcast segments).
Run C: configs/nanotrw512.yaml (66.5M parameters: width 512, 8 heads, batch expansion 4, frame budget 6000), trained from scratch on the CER/DNSMOS-filtered clean cache (32k rows, Whisper-large-v3 CER <= 0.1, DNSMOS OVRL >= 2.8), one RTX 4090, 60k updates.
How to listen
monitor/prompt-<uid>.wav: the reference voice given to the model (a real validation recording of an unseen speaker, decoded through the DACVAE codec, so it also shows the codec's own quality ceiling).monitor/<step-folder>/NNN.wav: the model's synthesis of caseNNN— the text of another recording of the same speaker, in the prompt's voice.monitor/<step-folder>/results.jsonllists per case:text(target transcript),prompt_uid(which prompt WAV),hypothesis(Whisper-large-v3 transcript of the synthesis),wer,cer,speaker_similarity,duration_ratio.NNN.jsonholds the sampler settings and timings.- Folder name suffix
-gG-uU-nN-sS-dD-kK: guidance G, guidance applied while t < U, initial-noise scale N, sway S, duration scale D, Euler steps K (no suffix = guidance 2, 16 steps). monitor/cases.json: the 48 (prompt, target) cases;monitor.jsonl: one summary row per evaluated checkpoint/setting;train.jsonl: training/validation curves;config.json: the full training configuration.
WER/CER: faster-whisper large-v3 (Turkish), Turkish text normalization (numbers spelled out, İ/ı-aware lower-casing, punctuation removed). SIM: microsoft/wavlm-base-plus-sv cosine between synthesis and the codec-decoded prompt. Corpus WER/CER over 48 cases of 10 held-out speakers; the ASR floor on real codec-decoded speech is WER ≈ 0.05 / CER ≈ 0.016.
Results
Configuration
{
"latent_dim": 128,
"width": 512,
"depth": 12,
"heads": 8,
"text_depth": 4,
"patch_size": 1,
"ff_mult": 3,
"cond_dropout": 0.2,
"reference_encoder": "mlp",
"reference_pooling": "mean",
"reference_paths": "both",
"duration_features": "baseline",
"positions": "rope",
"qk_norm": true,
"text_attention": 4,
"prediction": "edm",
"text_layout": "joined",
"duration": "rule",
"ctc_layer": 8,
"adaln_rank": 64
}{
"steps": 60000,
"batch_size": 96,
"accumulation": 1,
"learning_rate": 0.0008,
"warmup": 2000,
"weight_decay": 0.01,
"optimizer": "muon",
"muon_momentum": 0.95,
"ema_decay": 0.9999,
"precision": "bf16",
"workers": 4,
"worker_threads": 1,
"prefetch_factor": 4,
"loader_start_method": "spawn",
"cuda_prefetch": true,
"checkpoint_every": 1000,
"validate_every": 1000,
"log_every": 100,
"grad_checkpoint": true,
"compile": false,
"seed": 42,
"flow_reduction": "frame",
"duration_weight": 0.1,
"speaker_balance": 0.0,
"diagnostics_every": 0,
"pairing": "within",
"prompt_fraction_min": 0.1,
"prompt_fraction_max": 0.6,
"prompt_dropout": 0.3,
"time_sampling": "logit_normal",
"batch_expansion": 4,
"keep_every": 5000,
"ctc_weight": 0.1,
"contrastive_weight": 0.2,
"contrastive_margin": 0.1,
"wandb_project": "dacvae-tts-tr"
}