tongflow-modal-sam-audio

Official TongFlow plugin. Text-prompted sound separation with SAM-Audio (Meta, facebook/sam-audio-large), running on a GPU via Modal.

Capabilities

• Noise reduction (denoise_audio) — isolate the clean speech from a noisy recording.
• Vocal separation (separate_audio_track) — pull the vocals out of a mix.
• Extract track (music-extract) — isolate any sound described in free text ("dog barking", "the piano in the background", "drums"), not just fixed stem names.
• Separate sound (separate-sound) — split the audio into a described target and the residual (everything else) in one pass, so the same node covers "extract this sound" and "remove this sound". Optional time spans mark where the target occurs, which also enables picking out one specific speaker.

Credentials

Add in TongFlow Settings (gear icon, top-right):

Key	Required	Notes
MODAL_TOKEN_ID	✅	Create at modal.com/settings/tokens.
MODAL_TOKEN_SECRET	✅	Paired with MODAL_TOKEN_ID.
HF_TOKEN	✅	The checkpoint is gated: request access on facebook/sam-audio-large, then use that account's token.

On first use the plugin deploys to your Modal account automatically and caches the build; weights are cached on a shared Modal volume.

Tuning (env, optional)

Env	Default	Notes
SAM_AUDIO_MODEL	facebook/sam-audio-large	Any facebook/sam-audio-* variant.
SAM_AUDIO_RERANK	4	Candidates re-ranked per call; higher = better/slower.
SAM_AUDIO_PREDICT_SPANS	1	Predict where the target sound occurs first.
SAM_AUDIO_DENOISE_PROMPT	clear speech	What denoise_audio keeps.
SAM_AUDIO_VOCALS_PROMPT	singing vocals	What separate_audio_track keeps.
SAM_AUDIO_MAX_SECONDS	300	Longest accepted input.