How can I detect AI-generated speech and voice clones?
The AI-generated Speech Detection model determines whether a voice or speech recording was generated by an AI model (text-to-speech, voice cloning), or if it is a genuine human recording.
How it works
You submit an audio file to the API with the ai_speech model. The model only analyzes the acoustic content of the audio. No metadata is used, and real recordings that have been compressed, denoised or transmitted over a phone line are not flagged as AI-generated.
The model detects speech produced by the main voice generators, such as ElevenLabs, OpenAI, PlayHT, Resemble AI, Murf, WellSaid, Microsoft Neural TTS or Google WaveNet. New generators are added as they emerge.
The API returns a score between 0 and 1, along with individual scores for each supported voice generator. Scores above 0.5 generally indicate AI-generated speech. For high-precision workflows such as voice-based KYC or fraud detection, you can use a higher threshold such as 0.8.
Typical use-cases include fraud prevention in call centers, voice-based identity verification, detection of voice-clone scams and impersonation, and moderation of podcasts or voice messages.
AI-generated music
To detect fully AI-generated music tracks (e.g. Suno, Udio), use the AI-generated Music Detection model with the ai_music model. Both models can be combined in a single request.
Limits, operations and availability
Audio files can be up to 20 megabytes. 1 operation is consumed for each 10 seconds of audio analyzed, for each model requested. So if you analyze a 37-second audio file for AI-generated speech, that will count as 4 operations. AI-generated Speech Detection is available on paid plans.