VoiceHub/dacvae-tts-tr-stage2-b

DACVAE-TTS Turkish stage 2 (from run B 40k, clean data)

Generated audio of every evaluated checkpoint of the training run tr-stage2-b (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).

Stage 2: warm start (--init-from) from run B's 40k checkpoint (VoiceHub/dacvae-tts-tr-nano-b-ke4), trained 20k more updates on the CER/DNSMOS-filtered clean cache (32k rows) at LR 4e-4, batch expansion 4, frame budget 7000, one RTX 4090.

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 case NNN — the text of another recording of the same speaker, in the prompt's voice. monitor/<step-folder>/results.jsonl lists 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.json holds 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

step	folder	WER	CER	SIM	duration ratio	guidance	steps
2500	monitor/step-0002500-g3-u1-n1-s-1-d1-k32	0.185	0.101	0.964	1.03	3.0	32
5000	monitor/step-0005000-g3-u1-n1-s-1-d1-k32	0.220	0.120	0.962	1.03	3.0	32
7500	monitor/step-0007500-g3-u1-n1-s-1-d1-k32	0.197	0.105	0.960	1.03	3.0	32
10000	monitor/step-0010000-g3-u1-n1-s-1-d1-k32	0.184	0.090	0.961	1.03	3.0	32
12500	monitor/step-0012500-g3-u1-n1-s-1-d1-k32	0.189	0.101	0.948	1.03	3.0	32
15000	monitor/step-0015000-g3-u1-n1-s-1-d1-k32	0.171	0.087	0.943	1.03	3.0	32
17500	monitor/step-0017500-g3-u1-n1-s-1-d1-k32	0.151	0.079	0.944	1.03	3.0	32
20000	monitor/step-0020000-g3-u1-n1-s-1-d1-k32	0.147	0.082	0.945	1.03	3.0	32
20000	monitor/step-0020000-g4-u1-n1-s-1-d1-k32	0.140	0.073	0.944	1.03	4.0	32
20000	monitor/step-0020000-g5-u1-n1-s-1-d1-k32	0.144	0.070	0.946	1.03	5.0	32
20000	monitor/step-0020000-g6-u1-n1-s-1-d1-k32	0.127	0.063	0.943	1.03	6.0	32

Configuration

{
  "latent_dim": 128,
  "width": 448,
  "depth": 12,
  "heads": 7,
  "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": 20000,
  "batch_size": 96,
  "accumulation": 1,
  "learning_rate": 0.0004,
  "warmup": 500,
  "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": 2500,
  "ctc_weight": 0.1,
  "contrastive_weight": 0.2,
  "contrastive_margin": 0.1,
  "wandb_project": "dacvae-tts-tr"
}