The model can synthesize understandable Turkish speech, but it may retain a foreign accent due to the base model's acoustic priors and the practical adaptation ceiling observed with Qwen3-TTS-0.6B.
This model is not production-ready and should not be used for impersonation, fraud, non-consensual voice cloning, or misleading synthetic media.
Status
Property
Value
Version
v0.1-experimental
Base model
Qwen/Qwen3-TTS-0.6B-Base
Language target
Turkish (tr)
Current quality
Understandable Turkish with foreign accent
Production-ready
No
Primary checkpoint
Stage 2 step 2000
Selection method
Perceptual listening quality and EOS stability
Usage
This adapter requires the Qwen3-TTS base model, which should be downloaded separately from the Qwen repository.
Do not re-upload the base model weights.
python
1import torch
2from peft import PeftModel
3from qwen_tts import Qwen3TTSModel
4from huggingface_hub import snapshot_download
56# Download base model from Qwen7base_dir = snapshot_download("Qwen/Qwen3-TTS-0.6B-Base")89# Download this adapter10adapter_dir = snapshot_download("hcfk/qwen3-tts-turkish", subfolder="adapter")1112# Load base model13tts = Qwen3TTSModel.from_pretrained(14 base_dir,15 device_map="cuda",16 dtype=torch.bfloat16
17)1819# Load Turkish LoRA adapter20tts.model.talker.model = PeftModel.from_pretrained(21 tts.model.talker.model,22 adapter_dir
23)2425tts.model.talker.model.eval()
Write numbers, dates, units, and abbreviations as Turkish words before inference.
Example:
Correct: Türkiye Cumhuriyeti bin dokuz yüz yirmi üç yılında kuruldu.
Avoid: Türkiye Cumhuriyeti 1923 yılında kuruldu.
The base model was not trained natively on Turkish text patterns, and raw digits may trigger non-Turkish phoneme behavior.
Training Data
Property
Value
Dataset
issai/Turkish_Speech_Corpus
Source
Hugging Face dataset repository
License
MIT, as shown on the dataset page
Size
~179K training utterances
Audio
Resampled to 24 kHz
Attribution
ISSAI / TurkicASR affiliated work
Training Method
Property
Value
Fine-tuning method
LoRA / PEFT
Base model
Qwen3-TTS-0.6B-Base
Final selected checkpoint
Stage 2 step 2000
Code predictor
Frozen in final successful runs
Final selection metric
Perceptual quality, not loss alone
The final selected 0.6B checkpoint was produced through staged LoRA adaptation and selected by perceptual listening quality and EOS stability.
Important Findings
F1 — Sub loss did not reliably predict audio quality
Lower sub loss did not necessarily correlate with better perceptual audio quality.
In some experiments, training the code predictor reduced sub loss, but degraded acoustic quality. Therefore, the final checkpoint was selected by listening tests and EOS stability, not by sub loss alone.
F2 — Code predictor training degraded perceptual quality
Experiments with cp_lr > 0 reduced sub loss in some cases, but produced worse audio.
Final conclusion:
code_predictor should remain frozen
cp_lr = 0
F3 — MLP LoRA helped only with early stopping
Attention + MLP LoRA improved quality during the early training window, but longer training degraded audio and introduced artifacts.
F4 — Stage 2 step 2000 was the best 0.6B checkpoint
Stage 2 continued from the best early attention+MLP checkpoint, froze MLP LoRA, and trained attention LoRA only at ultra-low learning rate.
The best perceptual result for Qwen3-TTS-0.6B was: Stage 2 step 2000.
F5 — Partial full fine-tuning did not beat LoRA
Partial full fine-tuning of the last 2 transformer layers was tested, but did not outperform the best Stage 2 LoRA checkpoint.
F6 — Qwen3-TTS-0.6B appears to have reached its practical Turkish adaptation ceiling
Remaining issues include:
foreign accent
C→K substitution
imperfect Turkish phoneme/prosody mapping
These issues were not fully resolved with LoRA or partial full fine-tuning on the 0.6B model.
