Releases·Model Releasefeatured
Tencent releases AuK-Flash, a distilled 4-step variant of its 1.5B speech generation and editing model
Tencent released AuK-Flash, the distilled variant of AuK, a 1.5B open-source foundation model unifying speech generation and editing via natural-language instructions. AuK-Flash runs 4-step inference without classifier-free guidance, achieving a 4.5x wall-clock speedup over the full model. Weights are on Hugging Face and ModelScope alongside code, an arXiv report, and a task cookbook, making a single model usable across TTS, editing, enhancement, and separation workflows.
100 engagement2 sourcesTue, Sep 8, 2026, 03:59 PM
AI Overview
- Tencent's AuK is a 1.5B-parameter foundation model trained on millions of hours of audio, covering TTS, editing, enhancement, and separation via one instruction interface
- AuK-Flash is a distilled variant with fast 4-step inference; weights are on Hugging Face and ModelScope alongside the AuK base model
- AuK supports lyric editing, emotion, timbre, de-accent, whisper conversion, speech separation, music separation, and target speaker extraction
- Zero-shot TTS uses reference audio; Instruct TTS generates speech from a voice description alone, with a Cookbook of instruction templates and CLI/Python examples
Related
Releases·Model Release
FireRedTeam releases FireRedAudio, a 9B-parameter audio language model
FireRedTeam released FireRedAudio, a general-purpose audio language model built on a shared 9B-parameter LLM with decoupled continuous representations for understanding and generation. The model is available on HuggingFace with code and demos, enabling tasks like speech recognition, synthesis, and audio understanding in one framework.
FireRedTeam released FireRedAudio, a general-purpose audio language model built on a shared 9B-parameter LLM with decoupled continuous representations for understanding and generation. The model is available on HuggingFace with code and demos, enabling tasks like speech recognition, synthesis, and audio understanding in one framework.
70 engagement1 sourceredditFri, Aug 21, 2026, 04:05 PM
Releases·Model Release
SupraLabs releases SupraTTS-0.1-Beta, a 29.6M-parameter Glow-TTS-based text-to-speech model for edge devices and CPU
SupraLabs released SupraTTS-0.1-Beta, a ~29.6M-parameter text-to-speech model built on the Glow-TTS architecture. The team reports slightly better audio quality than original Glow-TTS at the same size, targeting edge devices and CPU inference. The model is available on Hugging Face with roadmap items including zero-shot voice cloning.
SupraLabs released SupraTTS-0.1-Beta, a ~29.6M-parameter text-to-speech model built on the Glow-TTS architecture. The team reports slightly better audio quality than original Glow-TTS at the same size, targeting edge devices and CPU inference. The model is available on Hugging Face with roadmap items including zero-shot voice cloning.
2 engagement1 sourceredditSun, Sep 27, 2026, 03:23 PM
Developments·Paper
FireRedAudio introduces unified audio language model with decoupled representations
FireRedAudio is a general-purpose audio language model that uses decoupled continuous representations to unify understanding and generation tasks. It is built on a shared 9B-parameter LLM and claims to be the first publicly disclosed model of its kind. This matters for practitioners as it addresses key representation challenges in audio AI.
FireRedAudio is a general-purpose audio language model that uses decoupled continuous representations to unify understanding and generation tasks. It is built on a shared 9B-parameter LLM and claims to be the first publicly disclosed model of its kind. This matters for practitioners as it addresses key representation challenges in audio AI.
0 engagement1 sourcearxivTue, Aug 25, 2026, 07:37 AM
Developments·Paper
Tencent details Yuanbao search agent training framework in arXiv report
An arXiv technical report describes the training framework behind Tencent's Yuanbao search agent, which uses cross-domain hybrid online policy distillation (OPD) to achieve search specialization without sacrificing general-purpose capabilities. The approach aims to avoid the alignment tax typically incurred when optimizing language models for specialized search behaviors.
An arXiv technical report describes the training framework behind Tencent's Yuanbao search agent, which uses cross-domain hybrid online policy distillation (OPD) to achieve search specialization without sacrificing general-purpose capabilities. The approach aims to avoid the alignment tax typically incurred when optimizing language models for specialized search behaviors.
0 engagement1 sourcearxivMon, Aug 3, 2026, 12:01 PM
Developments·Paper
Meta Researchers Unveil Alignment-Free Text-Audiobox for Voice Dubbing and Full-Duplex Dialogue Synthesis
A new arXiv paper introduces Text-AB, a unified framework for high-quality voice dubbing and full-duplex dialogue synthesis. It builds on a Diffusion Transformer trained with a flow-matching objective and departs from Meta's Audiobox system along three dimensions. The work matters to practitioners building speech synthesis and dubbing pipelines by offering an alignment-free alternative to prior architectures.
A new arXiv paper introduces Text-AB, a unified framework for high-quality voice dubbing and full-duplex dialogue synthesis. It builds on a Diffusion Transformer trained with a flow-matching objective and departs from Meta's Audiobox system along three dimensions. The work matters to practitioners building speech synthesis and dubbing pipelines by offering an alignment-free alternative to prior architectures.
0 engagement1 sourcearxivThu, Sep 3, 2026, 03:30 PM