Check a song, an audio file or a voice message for AI: the bot reads Suno, Udio and ElevenLabs labels and estimates a synthesized voice in speech.

Send music, an audio file or a voice message to @Synth_ID_bot. There is nothing special to do: audio can be sent as it is, as audio, as a voice message or as a document.

Labels in the audio file

Music and voice generators sometimes write a marker into the file. The bot reads C2PA Content Credentials and the tags of the file: ID3 in MP3, the tags of M4A and MP4, and the INFO block of WAV. When a marker names a generator, the bot shows it: Suno, Udio, ElevenLabs, Stable Audio, Mubert, AIVA, Google Lyria, OpenAI and others. See supported generators.

An AI voice in speech

Labels disappear whenever a file is recompressed, so the bot also listens to the sound. A neural network estimates whether the speech is synthesized: a text-to-speech voice-over, a voice clone, an ElevenLabs voice.

It runs only when the file has no labels, and it answers only when it is at least 90 percent sure. The answer then says Likely an AI-generated voice and shows the confidence.
It listens to the beginning of the recording, about the first 4 seconds of speech.
It does not judge recordings shorter than 4 seconds: on short phrases the network often gets a real voice wrong, and the bot says so instead of guessing.
In the bot's own words it finds about half of synthesized speech and occasionally reacts to a real voice. It is an estimate, not a verdict.
It does not check music or singing.

(sends a voice message, 0:12)

Synth ID Detector🤖 Likely an AI-generated voice: the file has no labels, but the neural network hears signs of synthesized speech (97% confidence). This is an estimate by an open neural network, not a verdict: it finds synthesized speech (ElevenLabs, TTS voiceovers, voice clones) about half of the time and occasionally fires on a real voice. To check an ElevenLabs voice for sure, use their AI Speech Classifier.

What the bot cannot see

Unlabeled music: a Suno or Udio track that went through a converter, a screen recording or a messenger.
A synthesized voice that the network misses; it misses about half.
Singing, songs and instrumental music: the voice check is made for speech.

Questions

Can the bot tell that a song was made by Suno?

Only if the file still carries Suno's marker. Without it the bot cannot tell.

Why does a short voice message come back with a note about length?

The network is unreliable on very short phrases, so a recording under 4 seconds is not judged by voice. Send a longer one.

What should I use for an ElevenLabs voice?

The bot links ElevenLabs' own AI Speech Classifier, which is more reliable for their voices. The bot cannot call such services for you, so it points you to them.

Checking audio and voice