Doseedo Film Score: spot a video, play your score on AI instruments, mix it to picture

SOUND EFFECTS (SFX) from words with Stable Audio Open — SOUND DESIGN and FOLEY for game audio, apps and video: impacts, whooshes, door slams, footsteps, UI blips, creature and sci-fi sounds, ambiences ("rain on a tin roof", "busy café"), risers, drum one-shots. A one-shot hit ("punchy house kick", "snare with a short room", "a riser into the drop"), a loop on the tempo grid ("glitchy percussion"), or a texture ("vinyl crackle", "rain"). kind picks which (hit | loop | texture); omitted, the server reads it off the words. Don't put a tempo in a hit's prompt — tempo words make it write a loop. Wrong tool for a part with notes (doseedo_render_midi) or a piece of music (you write it; doseedo_render_score plays it). Async: job_id → doseedo_wait_task (typically 15–40 s) → audio URLs (pre-signed: curl -OJ <url>, no auth header, valid 24 h). Cancel: a render that has not reached the GPU yet is stopped and its credits refunded; one already rendering finishes on the GPU (its credits are spent), the job ends as canceled and the output is discarded.

MIDI TO AUDIO, the virtual instrument for agents — THE tool for "I wrote the part, now play it": an AI instrument model (orchestral strings, brass, woodwinds, piano, guitar, bass, drums, World) performs your exact notes, in your key, at your bpm, and returns audio takes (doseedo_generate_music takes no notes or key). Input ONE of: notes[] ({pitch, start_beats, duration_beats, velocity} — beats from bar 1, the same rows as doseedo_edit_session set_midi_notes), midi_url (a .mid ≤2 MiB: a doseedo result link or a public https URL), or session_id + track_id (that track's notes, read from the session). Required: instrument (an entry's `call` from doseedo_instruments, no key — stay inside its range) and bpm. The notes are sent AS GIVEN — no pitch snapping, no quantizing. One voice per part: a chord part is rendered one voice per output with voice_split:true (result.voices maps each voice to its files and notes). Takes come back best-first (takes[0] = the scorer's winner). Async: job_id → doseedo_wait_task (~1–3 min). Result: takes[{url, rank, score}] (pre-signed: curl -OJ, no auth header, 24 h), seed, duration, voices?, refine?, chain. A take may miss or change a note — read takes[].check (measured against your notes) or listen; offer the takes, never promise note-perfection. Free harmony check first: doseedo_analyze_midi. DRUMS (GM notes): drum_engine:"oneshot" = a sampler kit made for drum_genre (default for "electronic"/"808"), every hit as written — use it on ANY kit when every hit must land; the drum model PERFORMS a kit and can drop hits (takes[].fit.missing). "synth" = a one-shot patch (synth_role, synth_genre); result.oneshots = the samples. SOUND: a preset (timbre_preset), a recording's tone (timbre_audio_url = timbre transfer), or YOUR OWN instrument (lora_id from doseedo_train_instrument; result.lora.applied). Cancel: a render that has not reached the GPU yet is stopped and its credits refunded; one already rendering finishes on the GPU (its credits are spent),

Make a CUSTOM INSTRUMENT in your own sound: upload 30 s–3 min of ONE instrument playing (your own player, a rare or folk instrument, a prepared/invented instrument, a vintage synth) and doseedo trains a private adapter (LoRA) on its serving instrument model in about 1–2 minutes. After that, doseedo_render_midi {instrument: <the same instrument>, lora_id} plays ANY notes you write in that sound. Several takes help (up to 8 files; the first 3 min are used, ≥8 s required); a clean, solo, dry recording works best. instrument = the closest catalog instrument — it picks the model voice the adapter trains on (a shamisen trains as acoustic_guitar, a whistle as flute: doseedo_instruments lists each one's call.instrument). The notes are transcribed for you (pass notes to skip that). 20 credits. Result: lora_id, name, the windows it trained on, loss first→last, and `use` (the render args). Earlier ones: doseedo_list_jobs {kind:"train_instrument"}. Owner-only: no one else can render with it. Async: job_id → doseedo_wait_task (~1–2 min). Cancel: the plane has no stop for a running training — it runs to completion (credits spent), the job ends as canceled and its output is discarded. Only a job still queued at the plane is refunded by the plane's own failure refund.

FILM SCORING, orchestral arrangement, background music, a song or an instrumental YOU wrote: a WHOLE ARRANGEMENT in one call (a film cue, a string quartet, a big-band chart): every part played by its doseedo instrument, one stem per part from bar 1 at your tempo, plus a summed mix. Input ONE of: parts[] ({name, instrument, timbre_preset?, lora_id?, players?, notes[]} — doseedo_instruments calls), midi_url (multi-track SMF), or musicxml_url (notation played as marked, at concert pitch; what isn't: prepared.score.warnings). A file's parts are matched by name, then GM program; name any part with instruments {"<part name or index>": {instrument, timbre_preset?, players?}}. A chord on a one-line instrument is split into voices (one stem each). ★ SECTIONS: an instrument is ONE player; give a part players: N (2–8) and its line is played by N different players (each its own performance), summed into ONE section stem — that is how you write an orchestra: violins I / violins II / violas / cellos / basses as five parts with players 6 / 6 / 4 / 4 / 3, horns 2–4 in unison, a woodwind or brass line doubled; leave players out for a solo. The WHOLE orchestra in one call: ≤ 32 lines and ≤ 96 player rows (players: N = N rows), plain, sampled and World / LoRA instruments together (~1 min groups), ≤ 600 s. Drums where every hit must land: parts[].drum_engine "oneshot" (+ drum_genre). Synth parts: doseedo_render_midi. Timing: bpm and/or tempo_map ([{beat, bpm}]; doseedo_tempo_map lands film hits on downbeats). Result: parts[{name, instrument, url, fit, timbre, lora, players}] — fit.verdict / coverage = did it PLAY its notes (a failed part is re-rolled once; fit.rolls.missed = a roll not played as one), lora.applied = its adapter played — plus mix_url. Next: doseedo_mix {from_job}, then doseedo_session_from_stems {from_job, mix_job}. Async: job_id → doseedo_wait_task (~1–4 min). Cancel: a render that has not reached the GPU yet is stopped and its credits refunded; one already rendering f

MIXING and MASTERING: each part levelled, panned, high-passed, given its depth on a shared reverb, and mastered to a loudness target — a WAV AND `mixgraph`, the mix as a document you edit. After doseedo_render_score pass from_job (its parts and the instruments they play are the stems); or stems: [{name, audio_url, instrument?}] (result links work as-is). An orchestra (strings / winds / brass / timpani / harp…) is SEATED by default (film-orchestral: violins left, low strings right, brass behind, each seat at its depth in a 2.3 s hall, written dynamics riding the faders); other stems get the blind default. style: film-orchestral, film-scoring-stage, concert-hall, film-epic-hybrid, cathedral-choral, classical-chamber, cartoon-studio, golden-age-cartoon, jazz-combo, ambient, country-pop, metal-modern, 1950s-mono, 1960s-lcr-pop, 1970s-analog-rock, 1970s-dub, 1980s-pop-gated, 1990s-rock-dry, 1990s-hiphop, 2000s-loudness-rock, 2010s-edm, 2010s-trap, 2020s-pop, 2020s-hyperpop — an id or words ("dry 90s rock"). balance {part: dB} moves parts on top of the levelling (+3 brings a melody forward). Change the mix: edit the returned mixgraph — strips[] {stem, gain_db, pan, seat, chain, sends}, buses[] (verb = the hall), master.chain, automation [{target: "strips.<stem>.gain_db", points: [[seconds, dB]…]}] for rides (dB off the strip's gain_db: 0 = as set, -60 mutes) — and call again with mix_job (this job_id) + only your edits (or the whole mixgraph + the same stems): rendered as written, rails reported, no credits. Then doseedo_session_from_stems {from_job, mix_job: <this job_id>} → the DAW project at this mix's levels and pans. TO PICTURE: dialogue (the VO; the music ducks under each line), music_offset_s (bar 1 on the picture), length_s (cut + end fade). Async: job_id → doseedo_wait_task (~15–90 s). Cancel: the plane has no stop for a running mix — it runs to completion (credits spent), the job ends as canceled and its output is discarded. Only a job still queued at the plane

Build a native DAW PROJECT from stems: each stem on its own audio track from bar 1, optional MIDI tracks (the written parts) beside them, on your tempo (bpm or tempo_map) — Logic Pro (.logicx) or Ableton Live (.als), zipped. The deliverable after doseedo_render_score (pass from_job: its job_id and the stems, MIDI and tempo come from that job) or after any renders / separations you made (stems: [{name, audio_url}] — result links work as-is). After doseedo_mix add mix_job: each track at that mix's level and pan, the mix as a muted reference track. Async: job_id → doseedo_wait_task (~30–90 s) → file (pre-signed zip, curl -o session.zip <url>, 24 h). Cancel: the plane has no stop for a running session build — it runs to completion (credits spent), the job ends as canceled and its output is discarded. Only a job still queued at the plane is refunded by the plane's own failure refund.

SPOT A FILM: a video in, the music editor's cue sheet out — the first step of scoring to picture. Scene detection on EVERY frame, and it listens to the soundtrack (CPU, no credits): hard cuts frame-exact (dark, low-contrast footage included), fades, flashes and visual impacts; each shot's brightness, colour, warmth, contrast, motion and mood, with a keyframe JPEG per shot so you can SEE what it shows; where dialogue is (Silero VAD) and loud audio hits; then cues placed where the picture wants music and the dialogue lets it (1M1, 1M2 …: start/end SMPTE timecode, emotion, energy, tension arc, hit points ranked by strength, a suggested bpm that lands them, and tempo_map_input — pass it to doseedo_tempo_map as is). It sees light, movement and editing, not meaning: read the keyframes for what each scene is about. Input: video_key (doseedo_get_upload_url accepts mp4/mov/m4v/webm/mkv) or video_url (public https), ≤ 2 GB and ≤ 3 h. Optional start_timecode (the first frame's, e.g. 01:00:00:00; default the file's own), coverage (0.05–0.98, default 0.55: how much of the film gets music), reel (cue ids <reel>M<n>), keyframes (0–60). Async: job_id → doseedo_wait_task (~1 min for a 10-min 1080p reel). Next: doseedo_tempo_map per cue → write the notes → doseedo_render_score → doseedo_mix → doseedo_score_to_picture. Cancel: the plane has no stop for a running spotting pass — it runs to completion (credits spent), the job ends as canceled and its output is discarded. Only a job still queued at the plane is refunded by the plane's own failure refund.

SCORE TO PICTURE: the last step of scoring a film. Lays each mixed cue (and any sound effect) on the video at its SMPTE timecode, with dialogue ducking (the lines found on the video's own soundtrack by VAD, or the spans you pass; the music dips ahead of each line and comes back after it), normalises the program to a delivery loudness (web −16, streaming −14, broadcast/ebu −23, atsc −24 LUFS, or a number; true peak under the ceiling) and muxes it onto the picture untouched (no re-encode). Returns the scored video, the printmaster WAV and the music-and-effects stem, plus a report of what it MEASURED: where each track landed (and what was cut at the film's edges), how far the music sits under the lines, and the final loudness and true peak. CPU, no credits. tracks: [{audio_url (a doseedo_mix mix_url as returned, or any public https URL) | audio_key, start_timecode (in the film's timecode) | start_s, gain_db?, role?: music|sfx, name?}], ≤ 64. Keep the production audio (default) or drop it (production_audio:"drop" — e.g. a temp score you are replacing; then pass dialogue spans to duck under). Up to 25 minutes of film per call (a reel). Async: job_id → doseedo_wait_task. Cancel: the plane has no stop for a running delivery — it runs to completion (credits spent), the job ends as canceled and its output is discarded. Only a job still queued at the plane is refunded by the plane's own failure refund.

MEASURE a take instead of guessing: give the audio (a render_midi take URL, any recording) and the notes it should play — notes[] + bpm (and tempo_map) or a midi_url — and get back how many written notes were heard at the right pitch and time (recall / precision / f1, octave errors, the first missed notes with their beats), the timing error in ms, loudness (integrated LUFS, peak dBFS, clipped share), any stretch where a note was written but nothing sounds, and a verdict (good / check / bad) with reasons in plain words. doseedo_render_midi results already carry this per take (takes[].check) when the part was given as notes. Method: an independent note transcriber (basic-pitch) — dense chords and very fast runs read less reliably; use it as a screen, and listen to anything marked check. Synchronous (~3–15 s), CPU, no credits. Cancel: not applicable — the result is returned in the submit response; the job is already finished.

Check the HARMONY of notes before rendering or placing them: factor a MIDI progression into its sonorities (mode "midi", default — chord-detection-free: each segment of simultaneously sounding pitches reduced to an anchor pitch class + register-exact shape, covered with recurring shape cells) or into chord symbols (mode "chord": one estimated chord per ~2 beats + a corpus-grammar cell cover with a schedule). Input: notes[] (the rows you would give doseedo_render_midi / set_midi_notes) with bpm, or midi_url (a .mid ≤2 MiB: a doseedo result link such as a stem's MIDI, or a public URL). Synchronous (~2–10 s), CPU, no credits. Returns sonorities[] or chords[], cover, cells, compression — read chords[].chord / sonorities[].anchor to confirm the key and progression you intended. Cancel: not applicable — the result is returned in the submit response; the job is already finished.

The instrument list — the sound palette: every instrument doseedo can PLAY YOUR NOTES on (doseedo_render_midi) — orchestral strings, brass and winds, guitars, basses, keys, drums and percussion, and World instruments (koto, sitar, banjo, pedal steel, tabla, congas…). Each entry gives the EXACT render call (`call`: instrument, plus timbre_preset where one is needed — a koto is the acoustic_guitar model with the koto timbre), its playable range in concert MIDI pitch, whether it plays chords, its presets (pizz, muted, sax registers…) and a demo render. No args → the whole catalog; query:"warm low brass" / family / pitches:[…] (instruments whose range holds them) / chords:true narrow it; id:"cello" → one instrument in full. Read-only, free, no key needed — read it before writing a part.

Scoring to picture: turn HIT POINTS (a cut, an impact, the title card — as seconds from the cue start or SMPTE timecodes) into a TEMPO MAP that lands each hit exactly on a downbeat (or a beat / an eighth — per hit with hits[].land), staying near your tempo and inside bpm_range. An AD or any cue that must hold ONE tempo: fixed_bpm: true (+ max_deviation_bpm, tolerance_ms) finds the tempo and the bar-1 offset (music_offset_s) landing the most hits, and says which ones it could not. Returns tempo_map [{beat, bpm}] to pass unchanged to doseedo_render_midi / doseedo_render_score / doseedo_session_from_stems, and hits[] with the beat, bar and beat-in-bar each hit lands on — write the accent, stinger or change there. Handles 23.976 / 24 / 25 / 29.97 (";" = drop-frame) / 30 / 50 / 59.94 / 60 fps timecode. Pure math: free, instant, no key needed.

Get a presigned upload for your own audio file — or a SCORE (.mid / .midi / .musicxml / .xml / .mxl → a score_key for doseedo_render_score: send a big score as a file, not as notes in the call). Returns an upload_url (PUT the raw file bytes there, e.g. `curl -X PUT -H "Content-Type: <content_type>" --data-binary @song.wav "<upload_url>"`) and an audio_key to pass to doseedo_separate_stems / doseedo_transcribe / doseedo_cover_song / doseedo_audio_to_session, or to doseedo_add_audio_to_session. A FILM (mp4, mov, m4v, webm, mkv; ≤ 2 GB) uploads the same way: pass the key as video_key to doseedo_spot_video / doseedo_score_to_picture. Formats: WAV, AIFF, FLAC, MP3, M4A/AAC, Ogg/Opus, WebM — all accepted for processing on every plan. Storing audio IN A SESSION (doseedo_add_audio_to_session) keeps WAV/AIFF/FLAC lossless only on Pro+; below Pro+ lossless is not kept there, so send Opus (audio/ogg). Samples/regions for a session that are on this machine need NO upload: use load_quick_sampler_sample sample_path / attach_audio local_path / set_track_clips audio_ref.local_path and run place_local_files after the download.

Check an async doseedo task once. Returns current status, and the result (file URLs / notes) when finished. A big result comes back as a brief (each row cut to its deciding fields): parts: ["<name>" or index, …] returns those rows of the job record IN FULL — a render_score part's every receipt, a mix part's strip, rides (every point) and loudness.

Wait for an async doseedo job to finish: polls server-side for up to timeout_seconds (5–50; one call can never hold longer than 50 s), then returns the result or the latest status. If it returns still_running:true, call it again — a submit reply's eta_seconds says how long the whole job usually takes (transcribe ~15 s, a render 1–3 min, audio_to_session ~4 min), so a 3-minute job is three or four calls, not one. A finished job answers with its output, artifacts (pre-signed links) and note.

List the jobs this account submitted (generate / render_midi / cover / stems / transcribe / session / convert / bounce / the analyses), newest first — to find an earlier result, check what is still running, or resume after a lost connection. Each row names its artifacts (e.g. ["vocals", "drums", "midi_piano"]); doseedo_check_task {job_id} returns the full result with fresh links. Read-only.

Cancel a queued or running job. What that stops and refunds depends on the kind — generate_music / render_midi / cover_song: a render not yet on the GPU is stopped and its credits refunded; one already rendering finishes (credits spent), the job ends as canceled, output discarded. audio_to_session: cooperative — stops at the next stage boundary and refunds the build's sound-tool unit (a separation already run stays charged). bounce_session: a bounce still queued never renders (budget refunded); one mid-render has its result dropped. separate_stems / transcribe / convert_project: no plane-side stop — the work completes (credits spent), the job ends as canceled and its output is discarded. The synchronous analyses are already finished (409 already_terminal).

Check the connected doseedo account: plan tier, the monthly credit pool (generation AND DAW conversions draw it), purchased extra-usage credits, and storage. Read-only — consumes nothing. Use it before starting expensive work to confirm there is budget.

无需输入参数。

Estimate what a job will cost in credits BEFORE submitting it, and whether the connected account can afford it right now. Spends nothing. Pass the job (generate_music, generate_sound, cover_stem, fill_gap, render_midi, train_instrument, cover_song, separate_stems, transcribe, audio_to_session, convert_project, bounce_session, spot_video, score_to_picture, analyze_song, detect_chords, beat_grid, analyze_midi, render_score, mix, session_from_stems, check_take) and the same args you would give the job tool (duration_seconds, models, instruments; for render_midi: notes + bpm / tempo_map or duration_seconds, voice_split / voices, takes — takes are free, voices cost one batch per ~6; for render_score: parts + bpm / tempo_map — planned exactly as the job plans it (lines, player rows, length), and `refused` says when the job would refuse it). The estimate follows the planes' own price table; the gate is authoritative and a refusal names the exact price.

The named pipelines people actually ask for — stems, session (mix → DAW project), cover, transcribe, generate, render-midi (your notes played by an instrument model), custom-instrument (train your own from a recording), timbre-transfer (your notes, a recording's tone), sound-design (SFX / foley / one-shots), analyze-song (key/tempo/sections/chords), convert, bounce, and the compose-part guide (a part in the session's key at its tempo: analyze → you write the MIDI → place or render it) — each with its inputs, the tool calls in order, outputs, a credit/time estimate and how to VERIFY success from the response. Call with no name for the index (pick by use_when), or with a name for the full recipe. Read-only.