voice_audio_multi/docs/models/midashenglm_gen.md
Fabio c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
Initial import: audio.cpp with Orange Pi documentation
2026-11-04 19:05:22 +01:00

2.4 KiB

MiDashengLM-Gen

MiDashengLM-Gen generates mixed audio from structured text prompts. Prompts can describe speech, music, sound effects, and environment layers using the upstream caption tags.

Quick Start

audiocpp_cli \
  --task gen \
  --family midashenglm_gen \
  --model models/MiDashengLM-Gen-GGUF/midashenglm-gen-f32.gguf \
  --backend cuda \
  --text "<|caption|> A calm narrator speaks over soft rain and a quiet room tone. <|asr|> The lights are still on, and the rain is getting softer now. <|speech|> calm warm male narrator, close microphone, gentle emotion <|music|> sparse quiet piano chords <|sfx|> soft rain against a window <|env|> small quiet room at night" \
  --request-option duration_sec=10 \
  --request-option guidance_scale=2.0 \
  --request-option stop_threshold=0.5 \
  --request-option min_stop_step=5 \
  --request-option seed=20260817 \
  --out midasheng.wav

Model

Field Value
Family midashenglm_gen
Task gen
Mode offline
Packages midashenglm-gen-f32.gguf, midashenglm-gen-q8_0.gguf
Input Structured prompt through --text
Output Generated waveform

Useful prompt fields include <|caption|>, <|asr|>, <|speech|>, <|music|>, <|sfx|>, and <|env|>. Use <|unknown|> for an empty branch when you want the model to keep the tag structure but omit that content.

Common Options (use directly)

Option Values Default Meaning
--text structured prompt required Prompt with MiDashengLM tags.

Request Options (use with --request-option)

Option Values Default Meaning
duration_sec float model budget Target audio duration budget. Must be finite and positive; no fixed runtime upper cap.
guidance_scale float 2.0 Classifier-free guidance scale. Any finite value is accepted.
stop_threshold finite float 0.5 Stop probability threshold. Any finite value is accepted.
min_stop_step integer >= 1 5 Minimum generated steps before accepting stop; zero is rejected.
seed integer in [0, 4294967295] 0 Fixed unsigned 32-bit generation seed; zero is a fixed seed and negative values are rejected.

Session Options (use with --session-option)

Option Values Default Meaning
midashenglm_gen.weight_type native, f32, f16, bf16, q8_0 native Weight storage override for experiments.