The Recording Sessions
Google selected four professional performers for the voice types represented in the experiment: tenor Christian Joel, bass Frederick Tong, mezzo-soprano Joanna Gamble, and soprano Olivia Doutney. Google's project story says the singers worked alone in the studio for hours and described the supplied material as parts of opera written on one note, sung across a range of pitches. Together the sessions produced 16 hours of singing.
Labeling and Annotation
The public accounts mention a map of sound containing elements such as vowels, consonants, and different notes. They do not publish the files, labels, annotation process, or licensing terms as a reusable dataset. It is therefore accurate to say the recordings covered vocal material used for learning, but not to invent a detailed metadata schema.
Spectral Analysis and Feature Engineering
Audio machine-learning projects often transform recordings into representations that make useful patterns easier to learn. Spectrograms, pitch contours, and other acoustic features are common across the field. Google has not said which of those representations Blob Opera used, so they are useful background concepts rather than confirmed parts of this project's pipeline.
Training the Neural Network
David Li's published explanation says the 16 hours were fed into a convolutional neural network, which learned to reproduce each voice type based on what it heard. The official experiment page adds that listeners hear the model's understanding rather than the original voices. Training duration, hardware, loss functions, and input-output representation have not been publicly specified.
Fine-Tuning for Natural Expression
The final experience demonstrates that the trained system can connect pitch and vowel controls to a recognizable opera-inspired result. It does not reveal whether the team used adversarial training, perceptual losses, or a separate fine-tuning stage. Those methods belong in a general survey of audio generation unless a project source confirms their use.