FireRedAudio: Unified 9B Audio Model for 1-Hour Understanding & Generation
pmttyji · reddit · 2026-08-22
FireRedTeam released FireRedAudio, a 9B-parameter general-purpose audio language model featuring decoupled continuous representations. It uses a single backbone for both understanding (Audio Encoder) and generation (RedAE pathway), supporting ASR, audio understanding, zero-shot TTS, Instruct TTS, and speech editing. It can handle recordings up to one hour long with precise temporal grounding.
Also released is FireRedTTS3, a unified speech system with a Base variant for zero-shot cloning across 24 languages and 21 Chinese dialects, and an Instruct variant for natural language voice design and editing. The models report leading WER/CER and speaker similarity scores.
More from Multimodal
- fal launches Worlds for real-time creation and interaction with AI-generated worlds — Scobleizer · 2026-10-16
- Gummy worms as a physics test: Gemini 4 Argon vs. Claude Opus 5.5 elasticity sims — DynamicWebPaige · 2026-10-16
- 260M-parameter looped diffusion transformer beats 6x larger baselines with 5x less compute — burkov · 2026-10-16
- Seedance 2.5 demo: stretching one second of an explosion into a 30-second shot — LudovicCreator · 2026-10-16
- Octopus Short Film Made with Grok Imagine, Upscaled via Topaz — creatoroff · 2026-10-16
- Open-sourced 61-second computing history film generated entirely by Muse Spark — sophiamyang · 2026-10-16