OpenAI GPT-Live Audio Gets SynthID Watermarks Before EU AI Act

OpenAI has updated its GPT-Live voice model to include Google DeepMind's SynthID watermarking. This means all audio generated through ChatGPT Voice and the OpenAI API now has an invisible and inaudible watermark. What's more, OpenAI also launched a verification API. This allows any developer or organization to check if incoming audio files contain these provenance signals. This update arrived one day before new transparency obligations from the EU AI Act take effect. This change builds on OpenAI's earlier image watermarking rollout. For audio, SynthID embeds the watermark in the spectrogram, using psychoacoustic masking to keep it inaudible to humans. This method ensures the watermark survives various alterations, like MP3 compression or re-recording. The bottom line is that there's now a way to verify the origin of AI-generated audio, which could impact how we trust digital voice content.