FreyaTTS: A compact Turkish-first text-to-speech model introduced on arXiv
Researchers released a technical report on Freya-TTS, a 183.2M-parameter non-autoregressive flow-matching Diffusion Transformer for Turkish text-to-speech. The model operates in AudioVAE2's continuous latent space, enabling 48 kHz reconstruction without a tokenizer.
Entities
Related
Tokenizer Transplantation Fixes Autoregressive Collapse in Bengali ASR
Researchers propose a tokenizer transplantation pipeline to replace the English-centric byte-level tokenizer in Moonshine speech recognition models with the native-script BanglaBERT WordPiece vocabulary. This resolves catastrophic autoregressive collapse on Bengali, enabling edge-efficient ASR for morphologically rich non-Latin languages.
Paper identifies word-level control bottleneck in LLM-based TTS
A new paper highlights the inability of current LLM-based TTS systems to explicitly manipulate word-level acoustic attributes, limiting precise stylistic control and temporal alignment for applications like audiobook narration and video dubbing. The authors attribute this to scarcity of fine-grained annotated datasets and architectural challenges.
Researchers identify modality interference as root cause of full-duplex SLM degradation
A new paper on arXiv (July 7, 2026) presents a fine-grained analysis of optimization dynamics in full-duplex Spoken Language Models (SLMs), identifying severe modality interference as the root cause of knowledge degradation and compromised semantic integrity. The work aims to enable more natural and intelligent full-duplex SLMs.
KRONOS: Autoregressive latent diffusion for 3D molecule generation
Researchers introduce KRONOS, a latent autoregressive diffusion framework for 3D molecule generation that operates in the latent space of a pre-trained autoencoder. It addresses the limitation of diffusion models requiring molecular size a priori, supporting variable-length generation and context conditioning.
Soofi S 30B-A3B: Open-source MoE hybrid Mamba Transformer for German and English
Researchers released Soofi S 30B-A3B, a sovereign open-source Mixture-of-Experts foundation model for German and English. Its hybrid Mamba-Transformer design activates only 3B of 30B parameters per token, achieving throughput advantages for long-context deployment. Pretrained on 27 trillion tokens with up-weighted German data, it matches dense 14-27B models on English and German benchmarks while excelling in code tasks.