Qwen3-TTS — KSA Arabic Fine-tune
A fine-tuned version of Qwen/Qwen3-TTS-12Hz-1.7B-Base for Saudi Arabian (Khaleeji/KSA) Arabic speech synthesis.
Training data: vadimbelsky/KSA_Arabic_English_Dataset_13k — ~13 k Arabic utterances in the KSA dialect, filtered to 1–20 s duration.
How Arabic support was added
The base model ships with a fixed set of languages in its codec token vocabulary; Arabic was not among them. Adding it required changes at three levels:
1. Arabic language embedding — warm-start initialisation
Arabic was assigned codec token ID 2072. Rather than initialising this embedding randomly, it was set to the mean of all existing language embeddings before training:
ARABIC_LANG_ID = 2072
codec_emb = qwen3tts.model.talker.model.codec_embedding
existing_ids = [v for k, v in config.talker_config.codec_language_id.items() if k != 'arabic']
avg = codec_emb.weight[existing_ids].float().mean(0)
codec_emb.weight[ARABIC_LANG_ID] = avg
2. Language-conditioned codec prefix (4-token think block)
A 4-token block injects the explicit language ID through the codec channel:
pos 3: codec_think_id
pos 4: codec_think_bos_id
pos 5: lang_id ← Arabic token 2072
pos 6: codec_think_eos_id
pos 7: speaker embedding slot ← shifted +1 vs. base model
The sequence offset in the collator was adjusted from +8 to +9, and codec_embedding_mask[7] = False so the speaker embedding is injected directly from the speaker encoder.
3. Language auto-detection
dataset.py detects Arabic automatically from Unicode range \u0600–\u06FF, so no explicit language field is needed per sample.
4. KSA speaker registration
Speaker ID 3000 (ksa_speaker) was registered. The embedding is extracted from a reference KSA audio clip by the frozen speaker encoder and written directly into the safetensors weights — the saved model is fully self-contained.
Training setup
| Setting | Value |
|---|---|
| Base model | Qwen/Qwen3-TTS-12Hz-1.7B-Base |
| Training data | vadimbelsky/KSA_Arabic_English_Dataset_13k (Arabic subset) |
| Optimizer | AdamW, lr=2e-6, weight decay=0.01 |
| Precision | bf16 mixed precision |
| Gradient accumulation | 4 steps (effective batch ~32) |
| Gradient clipping | 1.0 |
| Epochs | 5 (this checkpoint: epoch 4) |
| Loss | talker_loss + 0.3 × sub_talker_loss |
All model parameters were fine-tuned (no LoRA). The speaker encoder was kept frozen during training.
Inference
Install dependencies:
pip install qwen-tts soundfile torch
Single utterance:
import torch
import soundfile as sf
from qwen_tts.inference.qwen3_tts_model import Qwen3TTSModel
tts = Qwen3TTSModel.from_pretrained(
"vadimbelsky/qwen3-TTS-KSA",
dtype=torch.bfloat16,
device_map="cuda:0",
attn_implementation="sdpa",
)
wavs, sr = tts.generate_custom_voice(
text="الحين سويت فنجال قهوة، توني صحيت من النوم",
speaker="ksa_speaker",
language="arabic",
)
sf.write("output.wav", wavs[0], sr)
CLI with infer_ksa.py:
python infer_ksa.py \
--checkpoint vadimbelsky/qwen3-TTS-KSA \
--text "وين تبي تلتقي الحين؟" \
--output out_ksa.wav
Output is 24 kHz mono WAV.
Supported speakers & languages
| Speaker name | Language |
|---|---|
ksa_speaker | arabic |
License
Apache 2.0 — same as the base model.