FreyaTTS Technical Report

2026-07-10 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

Freya-TTS is a compact, tokenizer-free text-to-speech model, primarily designed for Turkish, offering reliable and efficient conversational synthesis. This 183.2M-parameter non-autoregressive conditional flow-matching Diffusion Transformer (DiT) operates within AudioVAE2's continuous latent space, handling 16 kHz encoding and 48 kHz decoding. It advances TTS by enabling rule-free, end-to-end modeling from a 92-symbol Turkish character vocabulary, eliminating the need for phonemizers or discrete speech tokenizers. Freya-TTS also features non-autoregressive parallel denoising for simultaneous latent sequence prediction and a two-stage post-training recipe for enhanced speaker consistency and robustness on short inputs. On the Freya-TR-Eval benchmark, it achieves an 8.0% WER and 3.0% CER, surpassing larger open-source systems. Its real-time factor of 0.11 on consumer GPUs and faster-than-real-time performance on laptop CPUs make it suitable for resource-constrained edge deployments. The model weights, training/inference code, and evaluation benchmark are released under the Apache-2.0 license.

Key takeaway

For Machine Learning Engineers developing efficient, high-quality text-to-speech systems, particularly for Turkish, Freya-TTS offers a compelling solution. Its compact 183.2M-parameter Diffusion Transformer architecture and tokenizer-free approach simplify deployment and reduce computational overhead. You should consider integrating this Apache-2.0 licensed model for edge deployments or applications requiring robust, consistent speech synthesis from raw text, bypassing complex phonemization pipelines.

Key insights

Freya-TTS is a compact, tokenizer-free Diffusion Transformer achieving high-quality, efficient Turkish text-to-speech.

Principles

• End-to-end character-based TTS eliminates phonemizers.
• Non-autoregressive denoising enables parallel latent prediction.
• Targeted post-training improves speaker consistency.

Method

A two-stage post-training recipe involves single-speaker voice locking and short-utterance coverage to improve robustness.

In practice

• Deploy TTS on resource-constrained edge devices.
• Synthesize Turkish speech directly from characters.

Topics

• Freya-TTS
• Text-to-Speech
• Diffusion Transformer
• Turkish NLP
• Non-autoregressive Models
• Edge Deployment

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

Related on AIssential

• FreyaTTS Technical Report — FreyaTTS offers efficient, high-quality Turkish TTS via a compact, tokenizer-free, non-autoregressive Diffusion Transformer.
• BlueMagpie-TTS: A Token-Efficient Tokenizer, Language Model, and TTS for Taiwanese-Accent Code-Switching Speech — Tailoring text-side components like tokenizers and language models to specific linguistic contexts dramatically improves TTS performance.
• Neural Text-to-Speech for Myaamia: Speech Synthesis for an Indigenous Algonquian Language — Neural TTS is viable for low-resource Indigenous languages like Myaamia, offering a framework for revitalization while upholding data sovereignty.
• Staged Depth-Pruning Distillation of a Flow-Matching Text-to-Speech Teacher: A Compact Hindi Speech Synthesizer — Staged depth-pruning distillation enables compact, high-quality TTS models from large teachers, even with limited training data.
• The Higgs-tts-2-3b-base Model: A Text-to-Speech Foundation Model — The higgs-tts-2-3b-base model offers advanced, expressive TTS with emergent capabilities from a unified architecture.
• SAID: Accelerating Diffusion-Based Language Models via Scaffold-Aware Iterative Decoding — SAID accelerates DLLMs by prioritizing denoising on structural tokens and adaptively processing uncertain tokens.

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 Computation and Language.