voice_audio_multi/docs/models/stable_audio.md

Stable Audio

Stable Audio is wired as --family stable_audio --task gen. It generates music or sound effects from text, and the music model can condition generation on input audio for init-audio or inpainting workflows.

Stable Audio 3 Medium

Use the medium package for the larger Stable Audio 3 music model. It uses the same user-facing controls as the music path.

Field	Value
Family	stable_audio
Model directory	models/stable-audio-3-medium
Task	gen
Main mode	Text-to-music
Prompt language	en
Optional audio input	Supported by the Stable Audio request path when the model package provides the needed components

audiocpp_cli --task gen --family stable_audio --model models/stable-audio-3-medium --backend cuda --text "wide cinematic ambient music with soft piano and evolving strings" --duration-seconds 30 --out music_medium.wav

Stable Audio 3 Small Music

This is the default music-generation package. Use it for text-to-music, init-audio, and inpainting.

Field	Value
Family	stable_audio
Model directory	models/stable-audio-3-small-music
Task	gen
Main modes	Text-to-music, init-audio, inpainting
Prompt language	en
Optional audio input	--audio with audio_input_kind=init_audio or audio_input_kind=inpaint_audio

Text to music:

audiocpp_cli --task gen --family stable_audio --model models/stable-audio-3-small-music --backend cuda --text "uplifting house music with bright synths and festival drums" --duration-seconds 30 --out music.wav

Init audio:

audiocpp_cli --task gen --family stable_audio --model models/stable-audio-3-small-music --backend cuda --audio input.wav --text "turn this into mellow downtempo music with deeper bass" --duration-seconds 10 --request-option audio_input_kind=init_audio --request-option init_noise_level=0.45 --out init_audio.wav

Inpaint audio:

audiocpp_cli --task gen --family stable_audio --model models/stable-audio-3-small-music --backend cuda --audio input.wav --text "replace the masked sections with a tight snare fill and bright guitar swell" --duration-seconds 10 --request-option audio_input_kind=inpaint_audio --request-option inpaint_mask_start_seconds=2.5,6.5 --request-option inpaint_mask_end_seconds=3.5,8.0 --out inpaint.wav

Stable Audio 3 Small SFX

Use this package for short sound effects. It uses the same gen task but is intended for SFX prompts instead of music prompts.

Field	Value
Family	stable_audio
Model directory	models/stable-audio-3-small-sfx
Task	gen
Main mode	Text-to-sound-effects
Prompt language	en
Optional audio input	Not part of the basic SFX command

audiocpp_cli --task gen --family stable_audio --model models/stable-audio-3-small-sfx --backend cuda --text "footsteps on gravel, close perspective, crisp natural stone texture" --duration-seconds 8 --out sfx.wav

Stable Audio Foundation / Open v1

The same family also loads Foundation checkpoints using the Stable Audio Open v1 layout. Use --task gen --mode offline with the Foundation checkpoint path; the Stable Audio 3 packages above use a separate runtime path.

Foundation accepts these additional request options through --request-option key=value:

Option	Values	Default	Meaning
sampler_mode	pingpong, euler, dpmpp-2m, dpmpp-3m-sde	not set	Foundation sampler; legacy sampler remains accepted.
sigma_min	float > 0	0.01	Minimum noise level for the DPM++ samplers.
sigma_max	float > sigma_min	100	Maximum noise level for the DPM++ samplers; init audio uses init_noise_level instead.
rho	float > 0	1	Noise-schedule parameter for the DPM++ samplers.

The following common/request tables describe Stable Audio 3. Foundation also accepts the shared request controls, with duration capped by its checkpoint limit.

Common Options (use directly)

Option	Values	Default	Meaning
--text	prompt text	required	Music or sound-effect prompt.
--audio	WAV path	not set	Source audio for init-audio or inpainting.
--duration-seconds	seconds[,seconds...]	120	Target duration per prompt. Use one value for all prompts, or one comma-separated value per Stable Audio batch_size item.
--num-inference-steps	integer	8	RF diffusion steps.
--guidance-scale	float	1.0	Classifier-free guidance scale.
--seed	integer	random if omitted	Generation seed.

Request Options (use with --request-option)

Option	Values	Default	Meaning
prompt	text	not set	Fallback text prompt when text_input is absent; pipe separates batch prompts.
negative_prompt	text	not set	Negative prompt.
duration_sec	seconds or comma-separated seconds	120	Target duration per batch prompt; encode multiple durations as a comma-separated string, not a JSON array.
batch_size	integer	1	Prompt batch size.
num_inference_steps	integer >= 1	8	Diffusion steps.
guidance_scale	float	1	Classifier-free guidance scale.
apg_scale	float	1	Adaptive projected guidance scale.
sampler_mode	pingpong, euler	not set	Stable Audio 3 diffusion sampler. Legacy sampler remains accepted.
seed	integer >= 0	random	Sampling seed; omitted chooses a random seed.
duration_padding_sec	seconds	6	Extra generated padding before truncation.
truncate_output_to_duration	bool	true	Trim decoded audio to requested duration.
chunked_decode	bool	true	Decode the autoencoder in chunks.
init_noise_level	[0, 1]	1	Strength for audio-conditioned generation.
audio_input_kind	init_audio, inpaint_audio	init_audio	How the model uses input audio.
inpaint_mask_start_sec	comma-separated seconds	not set	Inpaint region start times.
inpaint_mask_end_sec	comma-separated seconds	not set	Inpaint region end times.

The corresponding legacy *_seconds request names remain accepted. Use only one name for each control in a request.

Session Options (use with --session-option)

Option	Values	Default	Meaning
stable_audio.max_batch	integer >= 1	1	Maximum prompt batch size.
stable_audio.weight_type	native, f32, f16, bf16, q8_0	not set	Weight storage type; defaults to native.
stable_audio.mem_saver	bool	false	Release staged graph/cache state after conditioner, diffusion, and autoencoder phases to reduce resident VRAM. Later requests may rebuild released graphs.

For backend weight-type controls, use audiocpp_cli --inspect --model <model-dir> --family stable_audio.