The Short Answer
Blob Opera is a browser-based machine-learning experiment in which four characters represent bass, tenor, mezzo-soprano, and soprano. You drag one blob up or down to change pitch and move it along the other axis to change vowel color. The other three voices respond and harmonize automatically. Google describes two trained models: one that learned to produce opera-inspired singing and one that learned to harmonize the player's input. The sound you hear is generated by the model, not a playback of the four singers' original recordings. That is the basic loop. The interesting part is everything hidden behind it: the source recordings, the two models, the real-time browser constraints, and the details Google has chosen not to publish.
Four Singers and Sixteen Hours of Source Material
Google credits four professional singers for the source material: tenor Christian Joel, bass Frederick Tong, mezzo-soprano Joanna Gamble, and soprano Olivia Doutney. Together they recorded 16 hours of singing. Google's project story describes the sessions as solo studio work in which the singers supplied operatic material written on one note and sung across a range of pitches. That kind of controlled material gives a model many examples of the same pitch with different vowels and voice types. It is also a reminder that the experiment began with real labor by real musicians, even though the finished blobs do not simply replay their voices. The recordings are not distributed as a public dataset, and the public materials do not include a dataset card, annotation guide, or licensing terms for reuse.
What the Singing Model Learned
David Li's public explanation says the 16 hours of singing were fed into a convolutional neural network that learned to reproduce each represented voice type. The official experiment page adds an important distinction: listeners hear the model's understanding of what opera singing sounds like, based on what it learned from the recordings. They do not hear the singers' original performances being stitched together. That distinction rules out a simple concatenative sampler. It does not tell us how the network represents pitch, vowels, transitions, or expression internally. Is the model generating a waveform directly, predicting an acoustic feature, or driving another synthesis stage? Google has not published that pipeline, so a confident diagram of the system would be speculation dressed up as fact.
The Second Model: Harmony in Real Time
Google says the team created two models: one taught the computer to sing and another taught it to harmonize. The second model is what makes the quartet feel responsive. When you move one blob, the other voices react around your input rather than remaining fixed. That is why a single gesture can produce a complete vocal texture instead of one lonely note. The public account stops at the behavior. It does not publish a chord vocabulary, a rule set, a training corpus, or a description of how far ahead the harmony model plans. We can hear that some results feel stable and others feel tense, but we cannot turn those impressions into a claim that the model follows a particular textbook of four-part writing. The honest description is behavioral: the model responds to the player's input in real time, and its internal method is not publicly documented.
Why the Quartet Feels Playable in a Browser
Interactive music only feels convincing when sound follows a gesture quickly and consistently. Delay can come from input handling, model computation, audio buffering, device drivers, and the output hardware. Blob Opera uses what Google calls modern web-audio technology and responds in real time, but the public materials do not specify the audio frame size, the sample generation method, or which computations happen locally versus through other browser resources. Google also notes that the experiment may not perform optimally on older devices. A separate, optimized 16-blob performance at Google I/O was described with WebAssembly SIMD, but that does not prove the public four-blob version uses the same architecture. The safe conclusion is that the experience is engineered for low-latency browser interaction, not that we know the exact optimization stack.
A Short History of Algorithmic Composition
Blob Opera belongs to a much older lineage of music made with systems. In 1787, a musical dice game attributed to Mozart let a player combine pre-composed fragments to produce a waltz. In 1957, Lejaren Hiller and Leonard Isaacson used the ILLIAC I computer to generate the Illiac Suite by encoding rules of counterpoint and harmony. Around the same period, Iannis Xenakis used mathematical probability to shape orchestral works. These projects differ enormously, but they share a question: how much of a piece can be specified by a process, and how much depends on a performer or a moment of choice? Blob Opera is not an autonomous composer. It is an interactive instrument in which a person supplies gestures and trained models supply voices and harmony. The history is useful context, not evidence about the experiment's training data or internal rules.
What Google Has Not Disclosed
Several details are often asserted online without a source. The public materials do not provide:
- The exact convolutional network architecture, layer count, or parameter count used for the singing model.
- Training duration, hardware, loss functions, or the precise representation the model predicts.
- The annotation scheme, dataset license, or full contents of the 16 hours of recordings.
- The harmony model's architecture, training corpus, chord vocabulary, or rule set.
- A named neural vocoder or a sample-level waveform-generation method.
- Audio buffer size, frame size, latency budget, or which inference work happens on the device.
- Quantization, pruning, distillation, or other optimization techniques used in the public version.
How the Two Models Meet Your Gesture
You do not need the internal details to hear the division of labor. Pick one voice and hold a note. The other three settle into a chord around it. Move slowly upward or downward and listen to how the harmony follows. Change the vowel without changing pitch and notice that the character of the sound changes even though the represented note stays put. Now lead with a different voice. The same gesture will produce a different balance because the input has moved to another part of the quartet. Keep in mind that screen proximity is not a documented harmony control. Moving two blobs closer together on the canvas may change the visual composition, but the musical response comes from the model, not from a hidden distance slider. Listen to the result rather than inventing a rule for it.
A Five-Minute Listening Exercise
Use this sequence to hear the system without trying to reverse-engineer it:
- Solo the bass and hold one note for five seconds. Listen to the lowest part of the texture.
- Add a slow slide upward and back. Notice whether the other voices move at the same time or respond after your gesture.
- Change only the vowel and describe the difference with three adjectives.
- Lead with the soprano and repeat the same shape. Compare the balance with the bass-led version.
- Record a thirty-second take and listen back. Name one moment where the quartet felt stable and one where it felt tense.
Blob Opera Is Not a Voice Simulator
The experiment represents four voice categories, not every human voice. It does not accept microphone input, teach breath support, display notation, or evaluate a singer. It produces vowel-based vocal sound rather than lyrics, and its harmony is automatic rather than a chord editor. Those limits are part of the design. A four-voice quartet with a forgiving interface is easier to understand than a complete vocal simulator, and the constraint is what makes the first minute so approachable. Treat Blob Opera as an interactive introduction to register, ensemble sound, and model-generated voice. It is a doorway into deeper listening, not a substitute for training or a scientific model of singing.
Frequently Asked Questions
These answers separate published project facts from details that are not documented.
Are the blobs singing recordings of real people?
No. Four singers recorded the source material, but Google says the experiment produces the model's understanding of opera singing rather than replaying the original voices.
How many models does Blob Opera use?
Google describes two: one for the singing voices and one for harmony. The second lets the other blobs respond to and harmonize the player's input in real time.
What kind of neural network does it use?
David Li publicly identified a convolutional neural network for the singing model. Google has not published the architecture or training recipe, and the harmony model's architecture is not documented.
Does Blob Opera use a neural vocoder?
That is not documented. Neural vocoders are a real technology used in audio generation, but applying the term to Blob Opera without a source would be a guess.
Can I play Blob Opera offline?
No. The experiment is hosted by Google and needs a connection to load. The controls may run locally after loading, but the page itself is not an offline download.
Why do the other blobs harmonize automatically?
A second trained model responds to your input. Google has not published the rule set or training data behind that response, so the audible behavior is the reliable evidence.
Is Blob Opera AI-generated music?
It is a collaborative AI instrument. The voices and harmony are model-generated, but a person shapes the performance through gestures. It is not an autonomous system that writes a complete piece from a prompt.
How long did training take?
Google has not disclosed the training duration, hardware, or loss functions. The public figure that is documented is the 16 hours of source singing.