Synthetic narration is now ordinary on YouTube explainers, TikTok and Reels — often openly, sometimes not, and occasionally in a cloned voice of someone real. You can check the voice in a video, but the detector works on audio files, so most of the job is getting a clean stretch of speech out of the video first.
TextSight analyses audio files — MP3, WAV, M4A, OGG or FLAC, up to 10 MB. It does not accept video files, and it does not fetch YouTube, TikTok or Instagram links. To check a voiceover you extract the audio track, trim it to a stretch of narration, and upload that.
Export the audio track from your video editor, or use ffmpeg. This writes the soundtrack as uncompressed WAV without re-encoding it twice:
ffmpeg -i video.mp4 -vn -acodec pcm_s16le narration.wavThen trim to 20–60 seconds of narration. The detector scores the first 60 seconds of a clip, so the part you upload should start where the voice does.
Use a download option the platform itself offers where one exists, or ask the uploader for the file — for a commissioned video, the source audio is the best evidence you can get. Respect the platform's terms of service on downloading. Screen-recording a video playing on your screen captures it after another round of compression and sometimes your room's noise too, which makes results less reliable; if you must, keep the volume steady and avoid speakers-to-microphone capture.
Videos are messier than voice notes. Background music, sound effects and fast cuts all sit on top of the narration, and the detector's model was built for speech, not music.
AI narration is a legitimate production choice, and many channels use it openly. Platforms including YouTube and TikTok run their own labels for realistic synthetic or altered content; their current policies are the reference for what must be disclosed there. A synthetic result on a faceless explainer channel is rarely news.
It matters more when the voice belongs to a real person who did not say the words — a public figure endorsing a product, a news clip, a colleague in an internal video. That is where you should combine a detector result with provenance: who first published the video, whether the person's official channels carry it, and whether the same speech exists in a trusted recording you could use as a reference clip.
The full routine: best copy, trim, check, read, verify.
Read the steps →Disclosure for your own AI narration, and what to do if your voice is cloned.
Read the guide →Where detection fits when a clip is about to be published.
Read the guide →Your first check is free with no signup — 3 a day with a free account. Your audio is processed and discarded, never stored.