A&R by listening:
sub-genre embeddings that find the next artist instead of the next chart.

The trending-list problem

If your discovery tool ranks by current streams, current playlists, or current social momentum, it will only find acts you already know about. This is true by construction, not by accident. Trending lists are aggregations of behaviour that is already happening — they cannot, in principle, surface an artist whose audience hasn't yet formed. The whole point of A&R is to find what no one has yet found, and a tool that ranks on observed behaviour is, by definition, the wrong tool for that job.

This isn't a critique of the existence of trending tools. They have their place — for marketing, for catalogue management, for A&R support tasks that are downstream of a signing decision rather than upstream. The mistake is when a label treats a trending tool as a substitute for the work of finding acts. The trending tool can tell you who is already breaking; it cannot tell you who will break.

Embeddings that listen

Our sub-genre embeddings are trained on the audio itself, not on social signals. The training objective is contrastive: tracks that are confirmed sonically related (as judged by working A&R teams, not by playlist co-occurrence or stream similarity) are pulled together in the embedding space; unrelated tracks are pushed apart. No streaming signal, no social signal, no platform engagement metric enters the loss function. Two acts are near each other in the embedding because they sound near each other to a careful listener — not because they share a playlist, not because their followers overlap, not because they got pitched to the same editor.

The result is an embedding space with a strange property: it surfaces acts who fit a sonic neighbourhood but have not yet built (or have not yet been afforded) the audience that the neighbourhood usually carries. Those are exactly the acts an A&R team is paid to find. The embedding isn't doing the A&R work; it's reducing the universe of candidates from hundreds of thousands of releases per quarter to a few hundred worth a longer listen. That reduction is where its value lives.

Calibration against working A&R

The honest question for any model that claims to support A&R is: did it surface, in advance, the acts that working A&R teams actually signed? We calibrated the embedding against signings made by twelve working A&R teams across 2023–2025 — a mix of major-distributed indies, independent labels, and management companies in the UK, US and Germany. For each signing, we asked: would the embedding have surfaced this act inside its top-2% recommendations for the team's roster brief, given the data available at the time of signing?

The recall figure is 71%. That means seven signings out of ten were inside what the model would have suggested. The remaining 29% were either acts the model didn't recognise the sonic neighbourhood of (typically traditional or hyperlocal acts outside the training distribution) or acts whose signing rationale was non-sonic — strategic catalogue moves, label-to-label relationship signings, signings driven by a manager rather than by audio. The 71% number does not measure success. It measures that the embedding is not blind to the kind of judgement A&R actually exercises.

The number is updated each quarter as new signings come in. We publish it because we think A&R tools should be defensibly evaluated, and because the only honest way to evaluate one is against the working judgement of people who do A&R for a living.

Used as a colleague, not an oracle

Nomad A&R is built to be argued with. Every recommendation comes with three audio references (here is what we think this is close to) and the implicit framing the model used (we surfaced this because the brief was scouting in the post-hyperpop / softcore sub-neighbourhood). The team can disagree with both the candidate and the framing. Disagreement is captured, fed back into a re-ranking layer specific to that team's taste, and improves the model's calibration to how the team actually hears.

The framing is important. Tools that surface recommendations without explaining the implicit frame become oracles — black boxes whose suggestions can only be accepted or rejected wholesale. A&R judgement doesn't work that way. The valuable conversation is "you surfaced this because you thought we were scouting here, but we're actually scouting there, so the framing is wrong and the candidate isn't what we want." That conversation only happens if the framing is visible. We make it visible.

Where this fails

Truly culture-defining acts are, by definition, outside any existing embedding space. The model will never find them. The work of recognising a culture-defining act — the work of hearing something that does not yet have a sonic neighbourhood — still belongs to humans. We are not pretending otherwise, and any A&R team that uses Nomad as a substitute for that work is misusing it.

The embedding also has known weak regions. Production techniques that fool the timbre channel are an active failure mode (we surface them where we can detect the fooling but we can't always). Lyric-dense genres in non-English languages are recovered worse than English-language pop, which is the regrettable consequence of the training data we were able to assemble at scale. Traditional music outside the training distribution — hyperlocal folk, regional traditional forms — is essentially invisible to the model and we are honest about that.

The teams using Nomad A&R productively are the ones who use it for what it's good at — narrowing a quarterly listening queue from many thousands to a few dozen, surfacing candidates inside a brief they've defined themselves, calibrating their re-ranking layer over six months to where they actually want it. The model is a colleague in that workflow, not an oracle. That's the only framing in which any A&R tool, ours or anyone else's, is worth the team's time.