Xiaomi releases MiDashengLM-Gen — a text-to-complex audio scene...

Xiaomi releases MiDashengLM-Gen — a text-to-complex audio scene generator

The Xiaomi team (xiaomi-research) has released the end-to-end model MiDashengLM-Gen, which generates audio scenes from structured text simultaneously — speech, music, sound effects, and acoustic environment.

Compiled by Sergey KostenchukPublished 2026-08-18Updated 2026-08-24

2026-08-18 Multimodal HuggingFace

🎧 Xiaomi releases MiDashengLM-Gen: from text — a full sound scene

The Xiaomi team has released an end-to-end model: from structured text it simultaneously generates speech, music, sound effects, and acoustic environment. The core is Qwen3-1.7B and DiT flow matching; output is mono WAV 16 kHz, 1–20 seconds.

🌍 It is the first open end-to-end generator of mixed audio scenes: a single model replaces a pipeline of separate TTS, music, and SFX models. On Seed-TTS, speech recognition error is reduced from 12.15% to 2.79%, narrowing the gap with Qwen3-TTS (1.24%).

👤 You can try it immediately: a demo in 7 categories — xingws.github.io/midashenglm-gen-demo, weights — mispeech/midashenglm-gen on Hugging Face. Limitations: 16 kHz, up to ~20 seconds, no voice cloning, 13.19% error in Russian versus 2.42% in English.

Source 1: https://arxiv.org/abs/2608.11804 Source 2: https://github.com/xiaomi-research/midashenglm-gen

Sources