DACVAE-Turkish

An experimental Turkish speech audio autoencoder fine-tuned from Aratako/Semantic-DACVAE-Japanese, which derives from Meta DACVAE. This release reconstructs audio; it is not a text-to-speech model.

The Turkish fine-tune used acoustic reconstruction, KL and waveform L1 losses. The semantic teacher and GAN were disabled. Although the starting checkpoint comes from a semantic fine-tune, this release does not claim newly trained or verified Turkish semantic alignment.

• Mono audio at 48,000 Hz.
• Continuous VAE latents: 128 channels, 25 frames/second (hop 1,920 samples).
• Checkpoint: waveform L1 weight 150, seed 42, update 1,000.
• Inference: FP32, posterior mean, watermark bypass.
• One training seed; preliminary development-set evidence, not a final quality benchmark.

Evaluation

All three models use the same protocol: FP32, posterior mean, watermark bypass, 48 kHz and −16 LUFS preprocessing. Reconstruction metrics use 256 fixed three-second validation center crops, spanning 18 episodes. ASR uses 32 full validation recordings totaling 224.24 seconds, with a fixed Turkish Whisper large-v3-turbo decoding/normalization protocol. These are different measurement units; ASR is not run against partial-crop transcripts.

Model	Mel ↓	STFT ↓	Mel + STFT ↓	Raw SI-SDR, dB ↑	WER, % ↓	CER, % ↓
Meta DACVAE	0.278626	1.064086	1.342712	16.8840	4.9892	2.5641
Japanese initialization	0.265590	0.971278	1.236868	17.4423	4.9892	2.5295
DACVAE-Turkish	0.226915	0.843268	1.070183	17.4190	4.1215	2.2869

Compared with the Japanese initialization, the combined spectral loss decreases 13.48%, while raw SI-SDR decreases 0.0233 dB. Compared with Meta DACVAE, combined spectral loss decreases 20.30% and raw SI-SDR increases 0.5350 dB. The Japanese initialization already improves on Meta, so the entire Meta-to-Turkish difference cannot be attributed to the Turkish training stage.

The Turkish ASR result is 19/461 word errors and 66/2,886 character errors; the Japanese result is 23/461 and 73/2,886. Differences occur in only two recordings. Original audio has 20/461 word errors and 59/2,886 character errors. A lower ASR word count than original audio is not proof of better audio quality.

Paired episode bootstrap for the Turkish-minus-Japanese SI-SDR difference gives a descriptive 95% interval of [−0.0824, +0.0369] dB. This is not a formal equivalence test. Validation was used repeatedly for development; intervals do not account for model selection or training-seed uncertainty. No untouched final test, blind listening/MOS, speaker-similarity or downstream TTS result is claimed.

Training and data

Initialization revision: Aratako/Semantic-DACVAE-Japanese@96adcf1937e1ff46ec0817a07c80f2d7d64998f0. Original Meta checkpoint revision: facebook/dacvae-watermarked@8680102d141858a21bd533543966a2eb2e569f92.

Data source: the private Turkish podcast dataset Vyvo/tr-dataset-12, revision 847aeb429875352823edbbd989bfc9792243a890. Its card reports 93.2 source hours. Filtering and exact-PCM deduplication retained 59.3641 training hours, 2.9163 validation hours and 2.4720 test hours. Splits are grouped by show/episode; physical-speaker disjointness is not established. Private audio, transcripts and record identities are not distributed with this model.

Available training data is not training exposure. This checkpoint received 2,000 three-second crops: 6,000 seconds / 1.667 hours of presented audio, not a full pass over the 59.36-hour training split. The wider local experiment series presented 8.833 hours across multiple models; that is not this model's training duration or unique data coverage.

Setting	Value
Training precision / device	FP32 / one RTX 5070 Ti
Updates / seed	1,000 / 42
Batch / gradient accumulation	1 / 2
Crop length	3 seconds
Optimizer	AdamW, LR 1e-5, betas (0.8, 0.99), weight decay 0
LR schedule	5-update warmup; gamma 0.99999
Trained modules	All active codec parameters; inactive watermark path bypassed
Posterior	Sampled in training; mean in evaluation
Objective weights	Mel 15, STFT 1, canonical Gaussian KL 1e-4, waveform L1 150
Teacher / adversarial loss	Disabled / disabled

The mel/STFT definitions are the project recipe, not a numerical reproduction of the original DAC or Japanese training objective. Waveform weights 0/15/150 were screened with matched 100-update runs, then 150 was followed for 1,000 updates from a fresh Japanese initialization. Earlier three-seed experiments used waveform weight 0; they do not establish seed robustness for this release.

The exported checkpoint was strictly reloaded; all 317 state tensors matched the training checkpoint exactly. See training_info.json and reload-verification.json for the bundled provenance. Export validation is not an additional quality evaluation.

Training source commit: ef31744fc12e40e9254f175e6920e78913a51107 in the training/report repository. That GitHub repository is private; access is required to inspect its full history. The standalone inference instructions below do not depend on that access.

Standalone inference

The model repository includes inference.py and requirements.txt. The private training GitHub repository is not required for inference. Use Python 3.11; Torch and torchaudio 2.8.0 are the tested versions. CUDA users should install matching Torch/torchaudio wheels for their system before installing the remaining requirements.

python -m pip install 'huggingface_hub>=0.34,<1'
python -c "from huggingface_hub import snapshot_download; snapshot_download('VoiceHub/DACVAE-Turkish', local_dir='DACVAE-Turkish')"
python -m pip install -r DACVAE-Turkish/requirements.txt

import sys
from pathlib import Path

import soundfile as sf
import torch

repo = Path("DACVAE-Turkish").resolve()
sys.path.insert(0, str(repo))
from inference import load_model, reconstruct

model = load_model(repo / "weights.pth", device="cuda" if torch.cuda.is_available() else "cpu")
audio, sample_rate = sf.read("input.wav", dtype="float32", always_2d=True)
audio = torch.from_numpy(audio.T)  # [channels, samples]
y = reconstruct(model, audio, sample_rate)  # [1, 1, samples_at_48_kHz]
sf.write("reconstruction.wav", y[0, 0].cpu().numpy(), model.sample_rate)

The helper converts to mono, resamples to 48 kHz, applies the evaluation loudness and peak policy, decodes the posterior mean in FP32 and trims codec padding. The output follows that normalization policy; input loudness is not preserved. Pin the downloaded HF revision when reproducing results.

Use the supplied watermark-bypass helper: setting decoder.alpha = 0 alone is insufficient in the pinned upstream implementation, because the acoustic output head must still execute. Watermark detectability/preservation has not been tested for this fine-tune. BF16 reconstruction produced substantial spectral-metric differences in local diagnosis, so FP32 is the recommended evaluated path. Long-recording streaming and drop-in compatibility with an existing TTS latent model have not been established.

Full-utterance evaluation covered 3–30-second recordings. The helper does not split long files automatically. In the local release smoke, warmed GPU calls matched the training implementation exactly; an initial-call difference also occurred when repeating the reference implementation itself. First-call/backend bitwise equivalence is not guaranteed; see inference-verification.json.

License and attribution

This release uses Apache-2.0; preserve the supplied license and attribution notices. The immediate Japanese checkpoint's pinned model card declares MIT. Meta's pinned model metadata and pinned code license identify Apache-2.0. These upstream attributions are retained; see NOTICE.md for the lineage and the upstream model-card wording discrepancy.

Acknowledgements: Chihiro Arata / Aratako for the Japanese checkpoint, Meta's DACVAE authors, and Descript's DAC authors. These projects do not endorse this Turkish fine-tune. The model license does not redistribute or grant access to the private training recordings.