midashenglm_gen.md

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.

Options

Option	Values	Default	Meaning
--text	structured prompt	required	Prompt with MiDashengLM tags.
--request-option duration_sec=<seconds>	float	model budget	Target audio duration budget.
--request-option guidance_scale=<f>	float	2.0	Classifier-free guidance scale; 1.0 disables CFG.
--request-option stop_threshold=<f>	0..1	0.5	Stop probability threshold.
--request-option min_stop_step=<n>	integer >= 0	5	Minimum generated steps before accepting stop.
--request-option seed=<n>	integer, -1 for random	0	Generation seed.
--session-option midashenglm_gen.weight_type=<type>	native, f32, f16, bf16, q8_0	native	Weight storage override for experiments.