Qwen3 0.6B Turkish

A Turkish text-to-speech adaptation of Qwen/Qwen3-TTS-12Hz-0.6B-Base, trained on roughly 2,277 hours of Turkish read speech from Kiraat. Current listening samples show clear, natural-sounding speech with strong audio quality.

Give it a clean 5–10 second Turkish reference recording to choose the voice. It does not require per-voice fine-tuning or a fixed voice list.

Samples

Each sample uses a different voice from dataset and reads a paragraph. The reference clips and sample text were not used for training.

Model

• Base model: Qwen3-TTS-12Hz-0.6B-Base at revision 5d83992
• Parameters: ~1.1B packaged (0.6B talker, plus speech tokenizer)
• Audio frame rate: 12 Hz
• Sample rate: 24 kHz
• Language: Turkish (Auto conditioning; no new language token)
• Voice control: Reference recording → speaker x-vector
• Output: Non-streaming

Only the talker was trained. The audio codec and speaker encoder are frozen from the base model.

Training data

Training used serdarcaglar/kiraat at revision 75b3297, after filtering for duration, speech ratio, speaker consistency and audio quality.

• Training clips: 1,388,236
• Audio: ~2,277 hours
• Speakers: 54
• Source recordings: 2,540
• Held-out evaluation: 128 clips each for dev and test

Text was kept in its spoken form, and recordings whose text overlaps the evaluation sets were removed. Separate readings of the same text were kept.

Training ran for one epoch (21,691 updates) with 64 clips per update. Dev loss went from 2.885 to 2.710 across the run.

Try it yourself

from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained("erkamk/qwen3-0.6b-tr")
wavs, sr = model.generate(
    text="Merhaba, size nasıl yardımcı olabilirim?",
    language="Auto",                # no explicit Turkish token
    reference_audio="reference.wav",  # clean 5-10 second Turkish recording
    x_vector_only_mode=True,
    non_streaming_mode=True,
)

Use a reference recording with one speaker, no music or noise, and normal speech level. Match the settings above; other generation modes were not trained or tested. Paragraph-length text works well — a few sentences per call is a good range. See Limitations before using prices, dates or English words as input.

Limitations

Numbers written as digits. Training kept spoken forms, so 1.299,90 TL is often garbled. Write prices and dates out in words before synthesis: bin iki yüz doksan dokuz lira doksan kuruş. This fixes most cases without retraining.

English words and brand names. WhatsApp, iPhone, and mixed Turkish-English sentences come out unstable. Code-switching is not supported in this release.

One epoch. Dev loss was still improving when the epoch ended, so the model is under-trained rather than converged. Longer training and broader data should help.

Pronunciation-level issues like the first two are exactly what reinforcement learning targets: generate candidates, score them with an automatic Turkish judge, and update on the good ones. Doing that with a well-designed reward is the next step for this model. Until then, prefer spoken-form input.

License and attribution

Weights are released under Apache 2.0, inherited from the Qwen base model. Changes from the base checkpoint are described in Model above.

Training data attribution:

Serdar I. Çağlar, Kiraat: a Turkish read-speech corpus for text-to-speech, Hugging Face, 2026. https://huggingface.co/datasets/serdarcaglar/kiraat

Kiraat is gated and lists its source channels in its own card. Keep that attribution with any redistribution. Its terms state that rights to the original recordings remain with the uploaders, so dataset access does not by itself clear a narrator's voice for reuse.

You are responsible for using voice cloning and generated speech lawfully and with appropriate consent, including any jurisdiction-specific rules on synthetic voices and disclosure.

• Base model: Qwen3-TTS
• Training data: Kiraat