How Google can find a song from a tune you hum

A recording and a person humming its melody are acoustically very different. Google’s published explanation of hum-to-search describes a model that maps both into comparable representations, placing versions of the same tune near one another. It does not require the hum to reproduce the instruments, singer or original key.[1]

This is an existing capability introduced in 2020, not a new October 2026 release. Its appeal remains unusually direct: it turns a fragment remembered as sound into a search query without first requiring a title or lyric.[3]

A small musical example explains the challenge

Imagine an original four-note pattern represented by the pitch numbers 60, 64, 67, 64. Move every note five semitones higher and the numbers become 65, 69, 72, 69. The absolute pitches change, but both sequences move up four semitones, up three and down three.

There is a similar timing issue. Notes beginning at 0, 0.5, 1.0 and 1.5 seconds have the same relative spacing as notes beginning at 0, 0.75, 1.5 and 2.25 seconds. One performance is slower; its pattern of evenly spaced entries remains. A recogniser needs to tolerate some changes without declaring every vaguely similar melody identical.

These tiny examples are deliberately incomplete. Four notes may occur in many pieces. Distinguishing a song requires enough of the sequence, rhythm and context to narrow the candidates. But the examples explain why comparing two raw audio waveforms sample by sample would be the wrong starting point.

Google learned a representation rather than requiring sheet music

The published system uses audio spectrograms and learned embeddings. Training draws matching hummed and recorded examples together while separating non-matches. Google describes using real hummed examples and generated approximations of humming to obtain enough paired material. The aim is to retain the tune while reducing sensitivity to how it was performed.[1]

SPICE, Google’s self-supervised pitch-estimation work, provided a way to estimate pitch without a manually labelled training example for every note. Google’s hum-to-search account uses pitch extraction in creating training material. That is different from saying that every live query must first be turned into an explicit MIDI score.[4][1]

Think of the learned representation as a search key rather than a transcript. A transcript tells you which words or notes were recognised. A search key needs to bring a query close to the right candidate. Those goals overlap, but a useful retrieval system does not have to expose a perfect written score along the way.

Try the recognisable part, not the whole song

Google’s documented mobile route is to open the Google app, tap the microphone and choose Search a song, then hum, whistle or sing. The launch guidance suggests roughly 10–15 seconds. The system presents possible matches, which you can open and listen to. Interface and availability can vary by device, language and region.[2][3]

For a first attempt, choose the section you remember most clearly and keep its rhythm steady. Do not add invented words merely to make the query sound more like singing. These are practical suggestions based on the matching problem, not the results of a listening trial conducted for this article.

The striking part of hum-to-search is the change in what counts as a match. A person supplies an imperfect performance of a melody; the system looks for a recording with a compatible learned representation. Recognising the song depends on preserving the right musical relationships across very different sounds.

Report an error or suggest a correction