Datasets:
audio audioduration (s) 6.12 10.7 | checkpoint stringclasses 1
value | text stringclasses 5
values | prompt stringclasses 5
values | whisper_transcript stringclasses 5
values | wer float64 0.05 0.32 | cer float64 0.03 0.18 | speaker_similarity float64 0.89 0.98 | dnsmos_ovrl float64 2.73 3.47 | seconds float64 6.12 10.7 |
|---|---|---|---|---|---|---|---|---|---|
step-0010000 | Bu sabah erkenden kalkıp sahilde uzun bir yürüyüş yaptım; deniz o kadar sakindi ki martıların sesi bile net duyuluyordu. | samples/prompts/prompt-00.wav | Bu sabah erkenden kalkıp sahilde uzun yürüyüş yaptım. Deniz o kadar sakindi ki martıların sesi bile net duyuluyordu. | 0.052632 | 0.03 | 0.971341 | 3.076535 | 6.12 | |
step-0010000 | Yapay zeka modelleri, büyük miktarda veriden örüntüler öğrenerek konuşma sentezi gibi karmaşık görevleri artık şaşırtıcı bir doğallıkla yerine getirebiliyor. | samples/prompts/prompt-01.wav | Yapay zeka modelleri büyük miktarda veri denen örüntü öğrenerek konuşma sentezi gibi karmaşık görevleri artık şaşırtıcı bir doğallıkla yerine getirebiliyor. | 0.157895 | 0.036496 | 0.963255 | 3.466278 | 10.68 | |
step-0010000 | Anneannem her bayram sabahı mutfakta baklava açar, evin içi tereyağı ve fıstık kokusuyla dolar, biz çocuklar da sofranın kurulmasını sabırsızlıkla beklerdik. | samples/prompts/prompt-02.wav | Anneannem her bayram sabahı mutfakta bakfakta baklava açar, evin içi tereyağı ve fıstık kokusuyla olar, biz çocuklarda sofranın kurulmasını sabırsızlıkla beklerdik. | 0.190476 | 0.067164 | 0.892063 | 2.725522 | 9.84 | |
step-0010000 | Toplantı yarın saat 14'te başlayacak; lütfen sunum dosyalarınızı öğleden önce paylaşın ve %20'lik bütçe artışı önerisini gündeme eklemeyi unutmayın. | samples/prompts/prompt-03.wav | Toplantı yarın saat 8.14'te başlayacak. Lütfen sunum dosyalarınızı öğleden önce paylaşın ve %20'lik bütbütçe artışı önerisini gündeme eklemeyi unutmayın. | 0.142857 | 0.094891 | 0.982427 | 3.229007 | 10.44 | |
step-0010000 | İstanbul'da akşam trafiği başlamadan Boğaz Köprüsü'nden geçmek istiyorsanız en geç dörtte yola çıkmalısınız, yoksa bir saatlik yol üçe katlanır. | samples/prompts/prompt-04.wav | İstanbul'da akşam trafiklenmeden, boğaz köprü sığmadan boğaz köpmek istiyorsanız en geç dörtte yola çıkmalısınız. Yoksa bir saatlik yol üçe katlanır. | 0.315789 | 0.180328 | 0.977337 | 3.068755 | 9.2 |
dacvae-tts Turkish 115M: trc-w640 (tr-combined + tr-dataset-12)
Checkpoints of the training run trc-w640, each with 5 outputs: Turkish zero-shot voice-cloning TTS
(dacvae-tts, branch turkish-tts; flow-matching DiT on frozen Meta DACVAE
latents, 48 kHz). Training in progress (update 61200 of 100000): every 10000 updates a new checkpoint and its 5 outputs are added here.
5 outputs per checkpoint
The same 5 unseen sentences, each spoken in the voice of a different held-out speaker (samples/prompts/, real
recordings never seen in training, decoded through the codec), with the demo settings: guidance 5, 32 Euler steps,
duration auto, one sample (no best-of-N). The dataset viewer above plays them (pick the checkpoint in the menu).
Raw model output: at guidance 5 it is loud (about -12 LUFS); the demo normalizes to -16 LUFS.
| checkpoint | Freya WER | Freya CER | Freya SIM | Freya DNSMOS | 5 outputs: WER | CER | SIM | DNSMOS | listen |
|---|---|---|---|---|---|---|---|---|---|
step-0010000 |
- | - | - | - | 17.2% | 8.3% | 0.957 | 3.11 | samples/step-0010000 |
step-0020000 |
6.93% | 4.05% | 0.957 | 2.87 | 5.1% | 4.0% | 0.961 | 3.21 | samples/step-0020000 |
step-0030000 |
- | - | - | - | 8.1% | 2.7% | 0.957 | 3.07 | samples/step-0030000 |
step-0040000 |
4.70% | 2.48% | 0.955 | 2.88 | 11.1% | 5.4% | 0.952 | 3.31 | samples/step-0040000 |
step-0050000 |
running | 7.1% | 3.7% | 0.949 | 3.17 | samples/step-0050000 | |||
step-0060000 |
- | - | - | - | 8.1% | 4.1% | 0.950 | 3.22 | samples/step-0060000 |
Freya = Freya-TR-Eval, all 495 sentences (3,911 words, 24 held-out voices), the published protocol (guidance 5,
32 steps, prompt-rate duration, one sample, faster-whisper large-v3 beam 5). It runs on the CPU next to the training
(about 2 hours per checkpoint), so a checkpoint gets it when the evaluator is free, newest first; the audio is not
uploaded, per-sentence transcripts are in freya/<step>/results.jsonl. WER is measured only on Freya-TR-Eval. The
published model VoiceHub/dacvae-tts-tr-w512 (66.5M) scores
WER 4.32% / CER 2.50%, SIM 0.946, DNSMOS 2.89 on it; FreyaTTS-183M 8.0% / 3.0%, XTTS-v2 11.1%. The 5 outputs are
for listening (their 99 words make one wrong word worth about 1 WER point).
step-0060000
| # | text | output | prompt voice | Whisper transcript of the output | WER | SIM | DNSMOS |
|---|---|---|---|---|---|---|---|
| 01 | Bu sabah erkenden kalkıp sahilde uzun bir yürüyüş yaptım; deniz o kadar sakindi ki martıların sesi bile net duyuluyordu. | 01.wav | prompt-00.wav | Bu sabah erkenden kalkıp sahilde uzun bir yürüyüş yaptım. Deniz Odar sakindi ki martıların sesi bile net duyuluyordu. | 0.105 | 0.985 | 3.07 |
| 02 | Yapay zeka modelleri, büyük miktarda veriden örüntüler öğrenerek konuşma sentezi gibi karmaşık görevleri artık şaşırtıcı bir doğallıkla yerine getirebiliyor. | 02.wav | prompt-01.wav | Yapay zeka modelleri büyük miktarda veriden örüntüler öğrenerek konuşma sentezi gibi karmaşık görevleri artık şaşırtıcı bir doğallıkla yerine getirebiliyor. | 0.000 | 0.939 | 3.46 |
| 03 | Anneannem her bayram sabahı mutfakta baklava açar, evin içi tereyağı ve fıstık kokusuyla dolar, biz çocuklar da sofranın kurulmasını sabırsızlıkla beklerdik. | 03.wav | prompt-02.wav | Anneannem her bayram sabahı mutfakta baklava açar, evin içi tereyağı ve fıstık kokusuyla dolar. Biz çocuklar da sofranın kurulmasını sabırsızlıkla beklerdik. | 0.000 | 0.867 | 3.16 |
| 04 | Toplantı yarın saat 14'te başlayacak; lütfen sunum dosyalarınızı öğleden önce paylaşın ve %20'lik bütçe artışı önerisini gündeme eklemeyi unutmayın. | 04.wav | prompt-03.wav | Toplantı yarın saat 14'te başlayacak. Lütfen sunum dosyalarınızı öğleden önce paylaşın ve %20'lik bütçe artışı önerisini gündeme eklemeyi unutmayın. Altyazı M.K. | 0.143 | 0.984 | 3.17 |
| 05 | İstanbul'da akşam trafiği başlamadan Boğaz Köprüsü'nden geçmek istiyorsanız en geç dörtte yola çıkmalısınız, yoksa bir saatlik yol üçe katlanır. | 05.wav | prompt-04.wav | İstanbul'da akşam trafiği başlamadan, Boğaz Köprüsü'nün başlamadan geçmek istiyorsanız en geç dörtte yola çıkmalısınız. Yoksa yoksa bir saatlik yol üçe katlanır. | 0.158 | 0.975 | 3.25 |
Model and data
Model: 640x14 generator (10 heads, low-rank AdaLN, LARoPE cross-attention,
CTC at block 9), 114.8M parameters; byte-level text with the turkish-v2
normalization (the serving frontend: numbers, dates, currencies, abbreviations and acronyms spelled out). Prompts:
70% another recording of the same speaker, otherwise cut from the clip itself. Muon, WSD learning
rate (0.0007, cooldown over the last 20%), 2x RTX 4090.
Data: Codyfederer/tr-combined (221,531 clips, 277 h) filtered by Whisper-turbo transcript agreement (CER <= 0.15 after removing fillers such as "eee"), DNSMOS OVRL >= 2.5, speaker-embedding consistency within each speaker label and speaking rate: 155,812 clips, 210.7 h, 1,904 speakers; plus the clean subset of Vyvo/tr-dataset-12 (32,226 clips, 72 h). Together 188,038 clips, 283 h, 2,561 speakers; validation/test speakers (tr-combined: whole videos) are never trained on.
Checkpoints
checkpoints/step-0010000.ptcheckpoints/step-0020000.ptcheckpoints/step-0030000.ptcheckpoints/step-0040000.ptcheckpoints/step-0050000.ptcheckpoints/step-0060000.pt
Each file holds the EMA weights used for synthesis (plus the raw weights). dacvae_tts/ is the package version that
reads them:
import sys
from huggingface_hub import hf_hub_download, snapshot_download
sys.path.insert(0, snapshot_download("VoiceHub/dacvae-tts-tr-trc-w640", repo_type="dataset", allow_patterns=["dacvae_tts/*"]))
from dacvae_tts.inference import Synthesizer
tts = Synthesizer(hf_hub_download("VoiceHub/dacvae-tts-tr-trc-w640", "checkpoints/step-0060000.pt", repo_type="dataset"), device="cuda")
tts.synthesize("Merhaba, bugün nasılsın?", "prompt.wav", reference_text="Prompt kaydının metni.",
output="out.wav", guidance=5.0, steps=32, duration_mode="auto")
train.jsonl holds the training and validation curves, config.json the full configuration.
- Downloads last month
- 314