audio being segmented outside of specified slider range (demo playground)

[aungkhant0911]

Hi, I have attached a sample audio below . Not sure if I am not doing right. The demo playground doesn't seem to work well when range-slider / segment slider is combined with text prompt. It just keeps segmenting outside of specified range.

For example, in the audio sample attached below, you can here "Wow that looks...." at around 5th second. There is no distracting noises or anything like that as you can here. However not matter how I try to isolate that particular portion with slider + text prompt, it doesn't seem to manage to isolate it. For prompt, I have tried 'woman', 'female', 'speaking', 'speaker', 'vocal' etc..

Audio_test.mp3

Thank you!

[xiejesses]

encountered the same issue and cannot be properly segmented

[chevalierNoir]

@aungkhant0911 the text + span should work as expected. Just tested the text prompt "wind blowing" and put span around 3-4s and the model correctly segments the wind noise out.

It not working well for your prompts is more the model issue. Our model works usually better in shorter audios (preferably ~10s). So when I try separating the first 10s of your example and prompt "speech", it works pretty well (see below).

isolated_sound (1).wav