How can I extract the speech from one speaker in an audio file?

Question

Assisting with a project and they need me to extract the speech from one speaker in an interview and make it its own audio file.

The audio I have are 1-1.5 hr interview recordings, and it is generally two speakers (interviewer and participant). I need to essentially get to a point where I have a single audio file of the participant's speech only. The recordings were done very casually in offices or similar spaces, so they are not the best quality and both speakers were recorded to the same channel. I do NOT have transcripts for these interviews - they have never been transcribed so it is really just the raw audio.

My first thought was to do this manually in Praat: Create a blank TextGrid for the audio file and make boundaries around the participant's speech only. However, this is very tedious and will likely take me 3 hours per interview, and given how many interviews I need to do this amount of time is not at all efficient.

I also tried using Otter.ai, but the text file output only gives the beginning timestamp for each line of speech, so I would need to manually insert end timestamps to get interval durations. I would also still need to use some program to parse the audio into the different speakers and extract the participant's speech. I am not sure how to do this given that very basic text file output. I know I can format it for ELAN, and ELAN can produce Praat TextGrids, but this process is still not time efficient. I am also not sure if the faculty I am assisting would be willing to pay for the subscription. (I'm not even sure what kind of format transcripts need to be in for forced aligners? Are there forced aligners that can assign different speakers?)

In short, I would love something that can:

1. Take the raw audio
2. A) Detect different speakers and parse the audio based on this so I can then combine only the participant's utterances into a single audio file, or B) Auto-transcribe the audio while also separating the speakers, then using some program (forced aligner?) to create a TextGrid that I can then work with.

I know this is really complicated and it may not be possible. So if there really truly is no possible way to avoid doing all this manually, please tell me so I know to stop researching different ways to achieve this!

Answer 1

A possible step in the right direction would be to look for pauses between speaker turns, using Praat annotate (to silences). This assumes that there is some pause between A and B speaking, and if they talk on top of each other, this won't work. You would still have to manually decide whether a given sounding interval was interviewer vs. interviewee and delete the "sounding" label for the one you don't want.