Burn word-timed captions onto a video (local path, URL, or PiAPI task_id).

Default flow transcribes the video's OWN audio with ElevenLabs Scribe (word timestamps), groups words into short chunks (max_words per caption), and burns styled overlays: white bold, black stroke, drop shadow, centered low-third, Hebrew rendered in correct RTL visual order (English/brand tokens stay LTR). Pass captions ([{text, start, end}]) to skip Scribe and burn exact chunks; style overrides (font_px, y_pct, fill, crf, font_path...). Run AFTER verify_generated_audio passes and BEFORE stitch_videos.

Delete a private asset.

Detect tempo and beat times of an audio track — pure analysis, no output.

Returns {"bpm": float, "beats": [seconds, ...]} with beat times ascending from the track start and covering the whole file. The tempo is reported at its lower octave, pinned into [min_bpm, max_bpm] (a track that reads as 172 BPM comes back as ~86) — pass the window to control which octave the grid lands on. Use this to beat-snap multi-clip seams onto a music bed before stitching.

Silence / no detectable pulse returns {"bpm": 0.0, "beats": []} (not an error).

Save one frame of a video (local path or URL) as PNG; default = LAST frame.

Use the last frame of clip N as image_first of clip N+1 (generate_seedance_first_last) to bridge multi-clip continuity. With upload=true the frame is also hosted on a temporary public URL so it can be passed as a reference directly.

Synthesize speech with ElevenLabs and write the audio to a temp file.

Generate a Kling Omni video (single prompt or multi-shot).

Compose a music track with Eleven Music (3-600s) from a text prompt.

For speech ads the bed must be COMPLEMENTARY, not the main event: prompt for minimal/low-key/background music ("soft, sparse, no melodic hook, low intensity"), keep force_instrumental=true so it never fights the voiceover, match duration_s to the video, then lay it under with mix_music_into_video (low gain + speech ducking).

Generate a Seedance video from a first (and optional last) frame image.

Generate a Seedance video; auto-chain Hebrew BVAC lipsync when language is Hebrew.

HEBREW (language="he") — this tool runs the ENTIRE BVAC chain itself; do NOT pre-build anything:

• Speech: synthesized from text (keep it Hebrew) with ElevenLabs eleven_v3 and voice_id — OR pass audio_path (local mp3) to use a pre-approved take and skip TTS. Audition flow: generate_elevenlabs_voiceover -> user approves -> pass its audio_path here. Never build a black carrier, upload audio, or register a carrier asset yourself; the tool does all of it and ignores yours.

• Prompt: pass ONLY the scene description, Latin-only (transliterate_hebrew first). The "@ImageN is ..." / "@Video1 ..." reference lines and the lip-sync mechanism are composed server-side — do not write them yourself (customize via image_roles / other_roles).

• romanized_text: ALWAYS supply your own PHONEMIC RESPELLING of text — spell the Hebrew the way an English reader sounds it out, because Seedance picks visemes from this text with an English-dominant classifier (linguistic romanization makes it mouth English: chazir -> "church"). Form: syllable- hyphenated, stressed syllable in CAPS, ' for schwa — חזיר -> khah-ZEER (NOT chazir); גבר, קום מהספה -> GEH-ver, koom meh-hah-SAH-pah; כושר -> KOH-sher. Every syllable sayable; gender/morphology respected but rendered by English sound; English/brand tokens byte-for-byte; sentence-final words with extra care (they drive the lip-sync hardest). A structural gate validates it; omit it and an LLM (OpenRouter-first) respells instead.

• task_type is forced to seedance-2-less-restriction: the lower-moderation tier required for asset-backed fictional personas. Requesting another type has no effect; this is intentional, not an error.

• Scribe QA gates run on the source audio, and on the generated video when wait=true.

References split into human_image_urls (faces/people) and other_image_urls (product/room/scene). Generated people MUST be private assets: human refs have to be asset://<id> (register them with upload_asset first) unless auto_upload_assets=true. image_urls is a legacy alias for human_image_urls.

Generate an ambient/diegetic sound-effect bed (0.5-30s) from a text prompt.

Generation only — the caller mixes the bed onto the video locally with its own ducking recipe. Describe the soundscape concretely ("busy gym ambience: muffled crowd murmur, low machine hum"), pass the clip/ad runtime as duration_seconds, and keep loop=true for a seamless bed.

Get a single asset's current state (refreshes its TTL).

Fetch the current state of a PiAPI task.

Host a local file on a temporary public URL (tmpfiles, ~1h retention).

Use for non-asset references that need a provider-fetchable URL: a local ElevenLabs mp3 going into audio_urls (English reference-audio lip-sync), a product/room photo for other_image_urls, etc. Human/persona refs do NOT go here — register those with upload_asset instead.

List private assets (optionally filter by status: active,processing,failed).

List available ElevenLabs voices.

Lay a music bed under a video's existing speech/ambience audio.

Gain is ADAPTIVE by default: both tracks are loudness-measured and the music sits music_below_speech_db LUFS below the speech (14 = clearly audible but secondary; 18-20 = barely-there). duck=true adds a gentle side-chain dip while someone talks. Pass music_gain_db only to force a fixed gain. Music for speech ads is complementary, not the main event. Video stream is copied untouched; inputs may be URLs.

Lay a voiceover as the PRIMARY audio over a (silent) video.

The VO plays at full level and the video stream is copied untouched. An optional bed_path (ambient/music) is mixed bed_below_voice_db LUFS under the VO with a gentle side-chain duck. This is the inverse of mix_music_into_video (which ducks a bed under speech ALREADY in the video). Output runs the video's length; the audio is padded with silence if shorter, trimmed if longer. ffprobe-verified.

Stretch/compress a clip to hit a target duration (or explicit speed).

Provide EITHER target_duration_s (speed is computed as source/target) OR speed (1.0 unchanged, 0.5 = half-speed/2x longer). Use this to slow a b-roll clip slightly to fill a voiceover segment without repeating footage. interpolate=true smooths slow-mo via motion interpolation (default off = frame duplication). speed is clamped to [0.5, 2.0]; outside that range is a ToolError. Audio (if any) is retimed too. ffprobe-verified.

Cut a master voiceover mp3 at split_points_s (seconds, strictly increasing).

Returns len(points)+1 ordered segments with paths and durations. Workflow for ads longer than one clip: generate ONE master VO with generate_elevenlabs_voiceover (with_timestamps=true), pick sentence-boundary timestamps, split here, then pass each segment's path as audio_path to its own generate_seedance_video call. Each segment must fit its clip duration.

Concatenate >= 2 clips (local paths or URLs, in order) into one MP4.

Hard cuts; every clip is normalized to the first clip's resolution/fps so mixed specs concat cleanly. Use this for the final assembly of a multi-clip ad — do not shell out to ffmpeg concat yourself.

Transliterate Hebrew text to Latin (via LMStudio/OpenRouter LLM) for visual prompts.

Frame-accurately cut a clip to an exact span (re-encode, not -c copy).

Provide EITHER duration_s (keep [0, duration_s]) OR start_s+end_s. Use this to cut a generated b-roll clip down to a voiceover segment's length. Resolution/fps/aspect are preserved; a silent clip stays silent. Returns the ffprobe-verified actual duration.

Register a private asset from a local path or public URL; return its asset://id.

Local paths are first uploaded to a temporary public host for ingestion (the asset persists on PiAPI ~days afterward, independent of that URL). If influencer_page (path to an Obsidian .md) is given, the asset id is recorded there under a managed section for reuse.

Run the generated-video Scribe QA gate on a finished Seedance task.

Use this after an async (wait=false) Hebrew lipsync job completes: pass the original spoken text (Hebrew) plus the task_id (or a direct video_url). Downloads the video, extracts its audio, transcribes with ElevenLabs Scribe, and judges against text (100% pass / 85-99% warning / <85% raises). Do not substitute local whisper or other ASR — this is the canonical gate.