kaanhgunay/pocket-tts-tr

Pocket TTS Turkish — v0.1 Base

Turkish 24-layer Pocket TTS teacher model based on Kyutai Pocket TTS.

This is the v0.1 base release, focused on high-quality Turkish speech synthesis. It is intended both for direct Turkish TTS inference and as a base checkpoint for further fine-tuning and depth distillation.

Model

• Architecture: Pocket TTS
• Language: Turkish
• Transformer depth: 24 layers
• Tokenizer: Turkish SentencePiece BPE
• Vocabulary size: 4,000
• Audio sample rate: 24 kHz
• Training checkpoint selected: 112,000 steps
• Generation temperature: 0.3

Training data

The model was trained primarily on:

serdarcaglar/turkish-tts-audiobooks

A leakage-controlled training/dev/test split and additional preprocessing, normalization, alignment validation and filtering were applied before training.

Usage

With Pocket TTS:

uvx pocket-tts generate \
  --config "hf://kaanhgunay/pocket-tts-tr/config.yaml@v0.1-base" \
  --voice "./voice.wav" \
  --text "Merhaba, bugün güzel bir gün." \
  --output-path output.wav

Replace HF_USERNAME with the repository owner's Hugging Face username.

Strengths

The model performs well on standard Turkish text, including:

• fluent Turkish sentences
• long-form sentences
• Turkish morphology and long agglutinative words
• numbers and dates when appropriately normalized
• punctuation and natural pauses
• voice-conditioned speech generation

Known limitations

Foreign-language words

This release is primarily trained for Turkish.

English and other foreign-language words may be pronounced according to Turkish orthographic and phonetic patterns. Code-switching is not considered a supported capability in v0.1.

A future release is planned to improve Turkish-English code-switching while preserving Turkish quality.

Beginning-of-generation artifact

A short transient or pronunciation artifact may occasionally be audible around the first syllable at the beginning of generated speech.

This behavior is being investigated separately from the acoustic/language quality of the model.

Versioning

• v0.1-base — Turkish 24L base teacher, 112k steps
• Future releases may add code-switch improvements and distilled variants.

Attribution

This model is derived from Kyutai Pocket TTS.

Training data attribution:

Serdar I. Çağlar, Turkish TTS Audiobooks: a 2,724-hour Turkish read-speech corpus for text-to-speech, Hugging Face, 2026.

License

Model weights are released under CC BY 4.0.

Users are responsible for ensuring that voice cloning and generated speech are used lawfully and with appropriate consent.