Lyria 3.5, Lyria and real-time music
Lyria in Google products is made up of music creation and live workflows. Record the name of the product, format, and duration of the audio clip used for testing. The first tested audio clip from the Lyria sample was detected by two processes used during testing.
Choose a documented source
Google lists both SynthID (for Lyria music) and NotebookLM (for podcasts creation). Save the original download link and provide the source. AI audio applications use other methods of audio provenance – don't expect that SynthID exists in every single sound made using AI.
For music, compare the arrangement
Pay attention to rhythmic variations, lower pitches, drum hits, stereo mixing and tone variations. It's easy to identify added distortions within silence sections. Compare loudness levels since more volume could give the illusion of better sound.
For speech, compare intelligibility
Listen carefully whether you hear any word changes, softened consonants, unnatural pauses or changes of the speaker's characteristics. An adequate test result would be one that keeps the message and is easy to perceive. Evaluate the entire recording rather than a small sample.
Run the baseline check first
Performing testing of the initial sample is needed with one of the supported detector services before any audio processing is applied. If no result was achieved, record this fact and choose another baseline sample. Then repeat testing of the final exported audio through the used service and record the audio format and duration.
Our current audio status
The mentioned audio processing technique is experimental and it isn't activated in the public image uploader. Verified audio support data will be published before audio removal service is introduced. Image testing and image subscription are not equal to audio removal assurance.