FreyaTTS Technical Report

2026-07-13 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, extended

Summary

FreyaTTS is a compact, tokenizer-free, Turkish-first text-to-speech (TTS) model with 183.2M parameters, designed for efficient conversational synthesis. It utilizes a non-autoregressive conditional flow-matching Diffusion Transformer (DiT) operating in the frozen continuous-latent space of AudioVAE2, which provides 48 kHz reconstruction from a 16 kHz encode. The model employs a rule-free, end-to-end approach using a 92-symbol Turkish character vocabulary, learning agglutinative morphology and number pronunciation directly from audio without a phonemizer. FreyaTTS avoids autoregressive error accumulation by denoising the entire latent sequence in parallel over a predicted duration. A two-stage post-training recipe, including a single-speaker voice lock, stabilizes speaker identity, reducing F0 standard deviation from 74.9 Hz to 5.0 Hz. On the Freya-TR-Eval benchmark, it achieves an 8.0% WER and 3.0% CER, outperforming larger open-source systems. It boasts a real-time factor of 0.11 on consumer GPUs and can run faster than real time on a laptop CPU, making it suitable for edge deployment. The model weights, code, and benchmark are released under Apache-2.0.

Key takeaway

For Machine Learning Engineers developing Turkish TTS solutions, FreyaTTS offers a compelling, resource-efficient alternative to larger multilingual systems. Its compact 183.2M-parameter design and non-autoregressive architecture ensure high-quality, reliable conversational synthesis, even on consumer GPUs or laptop CPUs. You should consider integrating this Apache-2.0 licensed model for robust, edge-deployable Turkish speech generation, especially where precise number and acronym pronunciation is critical.

Key insights

FreyaTTS offers efficient, high-quality Turkish TTS via a compact, tokenizer-free, non-autoregressive Diffusion Transformer.

Principles

• Decouple representation learning from generative modeling.
• Non-autoregressive generation prevents error accumulation.
• End-to-end character-level learning handles complex morphology.

Method

Train a conditional flow-matching DiT on a frozen AudioVAE2 latent space, then apply a two-stage post-training for voice lock and short-utterance coverage.

In practice

• Use a frozen, high-fidelity codec for efficient TTS.
• Implement post-training for speaker identity stabilization.
• Employ non-autoregressive design for robust number synthesis.

Topics

• Turkish Speech Synthesis
• Diffusion Transformers
• Non-Autoregressive TTS
• AudioVAE2 Codec
• Edge AI Deployment
• Voice Lock Post-training

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• FreyaTTS Technical Report — Freya-TTS is a compact, tokenizer-free Diffusion Transformer achieving high-quality, efficient Turkish text-to-speech.
• Tokenizer Transplantation: Mitigating Autoregressive Collapse in Edge-Efficient Bengali ASR — Replacing English-centric tokenizers with native-script vocabularies prevents autoregressive collapse in edge ASR for morphologically rich languages.
• How Mistral’s Voxtral TTS is Redefining Multilingual Voice Cloning with a Hybrid Autoregressive and Flow-Matching Architecture — Voxtral TTS uses a hybrid autoregressive and flow-matching architecture for high-fidelity multilingual voice cloning.
• Certas Palavras: A 1980s-90s Brazilian Radio Corpus to Test TTS Models in Noisy Multi-Speaker Dialogues — Certas Palavras offers a challenging Brazilian Portuguese dataset for robust TTS training in noisy, spontaneous environments.
• Improving quality and robustness in LLM-based text-to-speech systems — LLM-based TTS quality improves through targeted techniques addressing accent leakage, expressiveness, and reliability.
• A Practical Evaluation Method for Long-Form Simultaneous Speech-to-Speech Translation — A new evaluation method for long-form SimulS2ST uses ASR, forced alignment, and sentence embeddings to measure latency and quality.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.