Cut fillers, find clips, clean up speech
Jake Marsh ยท November 20, 2026
Three new audio models joined the lineup, bringing it to 34. Together they cover most of what an episode needs between recording and publishing.
Uhm finds every um, uh and hmm in a recording, timed to 20 ms, and returns a cut list your editor can apply. Send transcript: true and it also catches "like", "you know", repeated words and false starts, and hands back the transcript with them taken out. In our tests on the AMI meeting test set it scored an F1 of 0.68. It costs $0.0003 a minute, or $0.0012 with the transcript.
Clips reads a podcast, call or talk and ranks the 15 to 60 second moments worth turning into shorts, each with its times, a score and plain reasons like "Starts at the beginning of a thought". In our tests on AMI meetings, 75% of the top clip was content people had marked as key in their own summaries. Send audio for $0.0012 a minute, or your own transcript (SRT, WebVTT or Scribe JSON) for $0.0001 a minute.
Clear takes out fans, traffic, hum and kitchen noise, reduces room echo, and masters the result to podcast loudness. It scored a PESQ of 3.16 on the VoiceBank+DEMAND test set in our tests, up from 1.97 for the noisy input, for $0.001 a minute.
Try all three in the playground, or see the podcast recipe that chains them: about seven cents for a 30-minute episode.