Researchers have developed Freya-TTS, a compact, tokenizer-free text-to-speech model optimized for Turkish conversational synthesis. This 183.2M-parameter model utilizes a Diffusion Transformer architecture operating in a latent space, enabling efficient text-to-latent mapping and high-quality audio reconstruction. Freya-TTS achieves strong performance on the Freya-TR-Eval benchmark with low word and character error rates, outperforming larger systems while maintaining a fast real-time factor suitable for resource-constrained environments. The model, its code, and evaluation benchmark are released under the Apache-2.0 license. AI
IMPACT This release provides a highly efficient and compact Turkish text-to-speech model, potentially accelerating edge deployments and research in low-resource language synthesis.
RANK_REASON The cluster contains a technical report detailing a new open-source model release with associated code and benchmarks.
- Apache Software License 2.0
- AudioVAE2
- Diffusion Transformer
- Freya-TR-Eval
- FreyaTTS
- Hugging Face
- arXivLabs
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
