Hacker News Reader

Comment by craigcannon | Hacker News Reader

Hey HN!

Craig from YC here. This project is a follow-up to SpeechBoard, which was a text-based audio editor - https://news.ycombinator.com/item?id=15670827. Thanks for all the feedback there :)

We were surprised to find that many users just wanted transcripts, so Podscripter is an attempt to solve that.

Here's how it works: every time you publish an episode (or give us a file) we run it through a speech to text service. Then we split up the speakers by hand, which ends up being a fair bit of work and is why it's 24hrs instead of minutes . Then we email you the transcript.

Before I was podcasting at YC I had my own podcast and couldn't justify paying $1 a minute for transcripts. These machine generated transcripts get you most of the way there for a lot less money :)

Let me know what you think!

View on HN

Hey Craig,

I wrote a service, (nowhere close for public release), that segments audio based on speakers. You have to identify one speech segment, it is then capable of labelling others. It uses GMM and MFCC. Is something like this in the works? Cool idea! I consume a fair bit of podcasts, I can affirm that there is definitely a need for this

Nice! Let me know when it's ready :)

I've tried what's out there and still haven't found a solution that can consistency diarize well. So I'm doing some experiments on my end too.

Do try out https://scribie.com/transcription/free as well once. Our diarisation system is around 90% accurate on longer paragraphs and will be out this week.

For $10, I'd expect above 99% accuracy, especially since I can get an automated transcript for $5 for 60 minutes that presumably has a similar error rate to what you are offering.

Also, I'd expect it sooner than 24 hours, since I can get automated ones back in under an hour.

Not trying to be cold water. I am actually interested, but what sets you apart from the other cheaper automated solutions? Am I wrong about the error rate I can expect elsewhere?

Thanks for the feedback. Ours are diarized accurately. I've yet to see an automated service that does that well.

That's a valuable piece of info that I didn't see mentioned on the site. I'd add it if I were you.

Updated. Thanks for the feedback.

By way of context, human transcriptions cost about $1 to $1.50 per [EDIT: minute] --which is what I use for my podcasts. Accuracy is extremely good, especially if you flag obscure terms when you submit.

That's super cheap! Who do you use?

Uggh. Per minute. Sorry. So it costs me about $25-30 per podcast typically. Which is pretty reasonable but obviously adds up for longer transcriptions.

For a podcasts with good audio quality(which is usually a case for podcasts) you probably can use speech recognition tools that will cost you a fraction of this price. They are pretty good nowdays

They're really not--at least for my purposes. And $25-50 for an episode requiring minimal cleanup versus spending 30-60 minutes going back through the audio and fixing things up is a no-brainer for me.

ADDED: Machine transcriptions are pretty good for a lot of things such as search and quickly skimming content. But if you want something that people can read as an alternative to listening to the podcast, you pretty much have to use human transcription or budget a bunch of time to fix up.

Is there a way you can record then edit a podcast but keep track of which microphone different voices are coming from? Seems like you could make speaker identification easier that way?

Yeah, if you record on multiple tracks it's pretty easy but podcasting setups vary a ton.

What automated service do you use?

Google

Would this integrate directly with zencastr? They record tracks individually - and then mix them - so you'd have direct access to the individual tracks.

Seems like if you did that or made a way for people to upload their unmixed tracks, you could save some time on the whole thing?

Yeah, that's an awesome idea.

I had kicked it around with some potential customers but went with this simpler model just to see if anyone was interested.

Will most likely go for that next because you're right, diarization is insane :)

Hey Craig, I appreciate the hard work! Is there any way to demo this service for free, or any plans to in the future? The $10 isn't a big hit or anything, but I, as well as many others, I'm sure prefer to try a service before handing off their credit card info.

Yeah, we'll definitely have something free in the future. :)

For the launch we wanted to make sure we could handle all the first orders quickly.

I think this is super neat. The podcasts I am on do not make $10 an episode, so sadly, I can't justify it. But if I could, I would. I shared it with my podcasting friends.

Thanks!

Would you also do YouTube videos? There are many lectures online and it would be nice if more people transcribed them.

Yeah, we could definitely do videos.

I'm able to upload one WAV file per speaker, like most of the podcasters I would pressume. Would that make it easier to automate the split and this make the service cheaper?

I need to test this out more thoroughly. Is your speaker audio totally clean when they're not talking?

For all practical reasons I'd say yes, I'm using https://auphonic.com/ which does use machine learning to mute the parts where no human speaks, then they send this audio to third party (like yours) https://auphonic.com/blog/2016/12/02/make-podcasts-searchabl... then they get the files back so auphonic can bundle it for download with the rest.

So the real value add is when you have multiple speakers? Any suggestions on a reliable service that can handle a single speaker?

You wouldn't need a "service", you can just run it through a TTS program locally.

True. Although I assume you mean speech recognition and not TTS. Good recommendations there would be great :)

Oops, yeah, STT.

Can your service handle corrupted files? I record on the fly and I have some great interviews that I cannot play.

Are you using AWS Polly?

Google actually

Hello,

Just curious if you are able to speech-to-text phone calls? Or if any point you plan to do so.

Thanks

Reply on news.ycombinator.com