DJ mode

The player has two modes: NORMAL for ordinary playback and DJ for automatically mixed sets.

DJ is Soundsible’s built-in automatic DJ. It does not put a crossfade on a shuffled queue. It builds an editable musical route, analyses how each pair can meet, prepares the incoming track on a second deck, and performs the handoff at a chosen cue point.

DJ mode is currently labelled beta in the player. Its core mixing and route workflow is available now; the label sets the expectation that its planning and interface are still being refined.

Start a set

• With nothing playing, press Start a DJ session in the bottom player, then choose music. The DJ chooses an opening from that selection and starts the route there.
• With a track playing, open the player and switch from NORMAL to DJ. The current track becomes the first visible influence in Session.
• From a song’s menu in NORMAL or inside the shell, Start DJ from this song makes that song the session’s source. It is the next song the DJ mixes into, the way it mixes any other, and the music comes from it after that.
• While DJ is active, playing a song anywhere outside the player shell starts a fresh DJ session immediately from that song at 0:00. Its source and route replace the previous session, including all previous requests and influences. Playing a collection starts from its first song and uses the collection as the source. Session controls are inside the shell.
• In Session, use Mix with… to add an influence or Change… to choose a new direction using the same music browser.

A source is one piece of music

A song, an album, an artist, a playlist or the favourites is one source. Choosing a collection outside the shell starts a new DJ session from it. Inside the shell, its primary button moves the current session onto it while the playing song continues: the DJ draws the next songs from it. It never turns the collection into a list of requests. Asking for every song of a collection is still there, by name (Add to session) in the collection’s menu inside the shell, and those songs travel together: they belong to the session they were asked for in.

The three parts of the workspace have separate jobs:

What the DJ listens for

When a track can be analysed, Soundsible measures its tempo and beat grid, key, musical sections, intro and outro, and three magnitudes that can be compared between songs: notes per second, integrated loudness (EBU R128, in LUFS) and spectral flatness, which tells a pure tone (0) from noise (about 0.56). The DJ does not steer the session by any of them: the sources and the recommender choose the music, and the DJ mixes it. Depending on the material, the two-deck engine can perform:

• beatmatched long blends;
• bass swaps and filter blends;
• echo cuts and structural cuts;
• conservative fades when the analysis is missing or a more elaborate mix would sound worse.

The server chooses the technique automatically for each pair, with the same plan for web and Android. A beatmatched blend must fit the local tempo within the players’ speed limit and keep a steady shared pulse throughout the overlap. Whether a pair can be beatmatched is all that decides it: any two keys can be mixed. Key and tempo closeness choose which song comes next and which blend is used, never an echo cut or fade instead of a blend. Both songs move toward a shared tempo, each changing by the same amount and never more than 8 % (a turntable’s pitch range): 120 and 132 BPM meet at 125.7 BPM, each 4.8 % from its own tempo. The playing song reaches that tempo gradually in the seconds before the next one comes in, and the next song returns to its own tempo gradually after the blend. Without room for that ramp, only the incoming song moves, within the same range. Two songs share a pulse when whole numbers of their beats last the same: one to one, two to one, three to two (116 and 174 BPM) or four to three (90 and 120 BPM). Reading a song in two or in three is a choice of level — a 12/8 is a 4/4 with ternary beats — so the mix does not depend on it. The plan states which alignment it promises: the beats, or also the bars when both songs’ first beats are confirmed and their bars last the same. Even when only the beats are promised, the mix starts on each song’s estimated first beat. A recording without a steady pulse uses an echo cut or fade. A longer blend also needs compatible keys and enough measured quiet opening and tail in both songs, with a straight beat grid throughout; it never crosses an unmeasured gap to extend the mix.

When reliable pulses cannot be blended at the permitted speed, DJ can choose an echo cut. Its delay is one measured outgoing beat near the exit, corrected for the deck’s playback speed. An unreliable pulse uses a structural fade; missing analysis keeps a conservative fade.

For each refill, DJ chooses up to eight songs before looking at their analysis, then orders the measured songs for transition quality and continuity in notes per second, loudness, timbre and tempo. The scales come from that refill’s own measurements, with equal influence after scaling. Songs without analysis keep their shuffled positions, so already downloaded music cannot crowd them out. An empty session still opens inside the most recently added source.

The server prepares up to sixteen candidates per refill in the background, including previews that have not been downloaded yet. The choice follows the session shuffle, so a downloaded song gets no advantage in joining the route. Once the entire chosen pool is measured, DJ can insert a compatible bridge between two songs that would otherwise mix poorly, while keeping every original song. It also analyses the whole local music library in the background.

A newly acquired preview is analysed as soon as its complete audio is ready. Web and Android retry pending transition measurements before loading the next deck; an unavailable measurement keeps its conservative fade.

The engine limits tempo stretching and falls back safely instead of forcing two incompatible recordings together. Loudness levelling sits inside the same program path, so a transition does not need a sudden volume jump to feel alive.

Measured sections help place the mix: it can start during an outro or finish before a breakdown. A phrase starts on a bar where the arrangement measurably changes, a sound entering or leaving; between changes, phrases are counted in whole bars at the length the changes repeat, or in eight-bar steps that are marked as an assumption. A breakdown is where the kick and bass fall to less than half as loud, and a drop is where they come back after one; a passage that is only louder is neither. When the pulse is reliable and a measured phrase fits, the incoming song enters so its drop lands on a phrase start of the outgoing song at the end of the blend; an assumed phrase is used only when no measured one fits. Otherwise it keeps the ordinary intro. Both web and Android follow the same server plan.

The pulse and the first beat of each bar come from Beat This!, an open neural beat tracker (MIT licence) that runs locally with ONNX Runtime; its decisions are then placed on the attacks precisely enough to keep blended kicks together. The network can misplace a first beat by a whole beat or two, and its confidence does not show when, so a bar alignment is only promised where an independent measurement of harmony changes and accents agrees. Phrases and sections are counted on those bars, so a misplaced first beat moves them too; where the levels do not show a role clearly, the DJ claims none.

The model adds about 4 seconds of processor time per minute of audio, on one core, and up to about 0.5 GB of memory while a song is analysed. A first analysis of a large library therefore takes hours in the background; songs play with conservative fades until they are measured. Docker images and the desktop app include the model. An installation from source downloads it once (84 MB) the first time it analyses a song, and verifies its checksum. A recording whose tempo drifts, such as a live band, gets a grid that follows the performance and a fade at normal speed instead of a fixed-rate beatmatch. After an update that changes the analysis, the library and selected previews are measured again in the background; until those measurements land, their transitions are plain fades.

Whole songs, without mixing

Settings → DJ → Mix between songs is on by default. Switched off, the DJ keeps choosing the same music in the same route, but every song plays to the end of its file and the next starts from its beginning, with no fade, effect or tempo change between them. Next and Play now cut straight to their song, and Fix mix is hidden: there are no transitions to rebuild.

The switch is an account preference. It applies from the next song the DJ has not prepared yet; a mix already prepared still plays.

Make the set yours

• Add to session requests specific songs without changing the musical direction.
• Mix into session adds an influence alongside the existing ones.
• Inside the shell, Change session (and Start DJ from…) moves the session onto a new source at once. The old sources, the DJ’s own picks and bridges, and songs that were requested as a whole collection go with the old session; songs you requested one at a time stay. The current song continues and the DJ mixes out of it when it judges best; an already audible blend finishes. The new route is then planned like any refill, retrying on its own while the music plays on.
• The Session block shows the active influences. Remove an influence while another remains, or use Change… to replace the last one.
• Place a must-play track. Add a song to the Route or drop it into a particular gap. Soundsible may insert a bridge when that produces a safer path to the request.
• Reorder freely. Moving tracks changes the route immediately. Joins that no longer match their original transition use a plain fade until you press Fix mix.
• Repair without losing your choices. Fix mix rebuilds transitions and generated bridges around the tracks you placed while keeping those tracks in their chosen order and depth.
• Skip without leaving DJ. Next asks the DJ for a short handoff to the next route item instead of dropping back to ordinary playback. Next always does something: a blend that is already audible finishes on the spot, and a next song that cannot be blended in yet is started the ordinary way.
• Keep what the DJ finds. Every Route row has a visible ⋯ menu (also available by right-click or mobile long press). The menu saves the song to your library, downloads it, marks it as a favourite or adds it to a playlist without moving it in the route, including the song already cued.
• Choose music from anywhere. Song and collection menus offer the same session actions. Play now remains an explicit immediate handoff. Choosing a musical influence does not also request that exact song.

DJ keeps explicit requests ahead of generated music. Leaving DJ removes its generated branches and bridges, but tracks you explicitly placed survive as a normal manual queue.

The NORMAL / DJ badge in the mini-player always shows who owns playback. The choice is sticky for the current listening session and is included in session handoff state. Only selecting NORMAL explicitly, or accepting the prompt shown before starting incompatible Podcast or Radio playback, leaves DJ mode. If another song is selected while a blend is already audible, that blend finishes and the most recent selection is mixed next.

DJ, Radio, and Autoplay are different

DJ explores the active influences together with the last four automatic songs that actually started playing in the current direction. A pending recommendation is not a discovery root. Exact requests and bridges do not become roots just because they played. Mix with… retains this exploration; Change… clears it, without promoting the outgoing song. These discovery roots do not appear as additional influences in Session.

A change of direction takes effect at once: the new influences, a new exploration revision and the cleared route are committed together, and any answer still in flight for the old direction is discarded when it lands.

Every handoff ends somewhere

A transition is prepared on the second deck, and nothing fades until that deck is actually sounding. When it is not — the stream is still buffering, stalled, or ended seconds in — the handoff still resolves:

• if the outgoing song plays to its end first, the next song is handed over the ordinary way, from the deck that was already holding it, with the same start-up supervision and stall recovery as any other track change;
• a next song that cannot play at all is dropped and the one after it follows;
• a next song whose stream stops during an audible blend is given up and the outgoing song keeps playing, or, if it already owned the music, is recovered like any stalled song.

Recordings longer than twenty minutes — DJ sets, live takes, whole albums in one file — are not chosen as songs of a route unless a source holds nothing else.

If a provider request fails temporarily, DJ retries with increasing delays. If no new candidates remain, DJ stops repeating the same search and offers Retry. New influences, exploration roots or exclusions make planning eligible again. Existing playable route entries continue normally in either case.

Session snapshots preserve exploration separately from repeat history. Older snapshots keep their queue and requests, but history without reliable direction provenance is not promoted to exploration.

Continuity validation

A deterministic regression exercises thirty successive routes with overlapping recommendation neighbourhoods, plus independent tests for direction changes, request isolation, interrupted preparation and exhausted-input retry behaviour. A read-only local replay across 204 library tracks and cached related results produced 1,090 route extensions with both the repaired selector and the baseline at 2b5a365, compared with 972 at 34d95f1. Four tracks had no initial route in all three runs. Missing cache entries were treated as empty; no live provider requests or audio playback were performed. This checks selection continuity, not audible transition quality or physical-device playback.

Broadcast the result

Live captures the program bus after both decks, loudness levelling, EQ, filters, crossfade, and echo. Listeners therefore hear the same DJ set that the DJ hears, not a separate approximation. See Live for HTTPS and room setup.

Privacy and storage

Audio analysis is local. Soundsible stores compact measured features in its DJ cache, not decoded copies of the audio. Listening history and recommendation signals remain scoped to the account on your Station.

The detailed queue and route rules are recorded in Playback queue contract. Implementation and data flow are described in Architecture.