zlacker

I thought there were some open source speech-to-text models already [1].

Maybe there's something unique about how these low-quality radio transmissions sound that make these ineffective?

[1] https://voice.mozilla.org/en

>>ciaran+(OP)
I work in the speech recognition space and train my own models already. The existing open-source models aren't very good at noisy radio speech. I will specialize one of my models to this task once I have some data from the site.

>>lunixb+b
Got it, thanks. Good luck!

>>lunixb+b
As you’re well aware but HN folks may not be, it’s not just that it’s noisy, it’s heavily coded, contextually bankrupt speech between multiple parties that spend all day in contact with each other. Dispatchers in particular seem to have superhuman ability to extract information from completely unintelligible garbage.

Are you doing any kind of speaker identification?

>>jcims+u4
This is a very accurate description of the problem space. Every municipality has their own jargon, vernacular, and ways to communicate brevity which is key in public safety communications. The communications are often digitized over vocoders that are less than optimal, and then you have the process of recovering voice from noisy communcations channels.

This is definitely a very hard problem to solve.

>>blanto+fa
Indeed. The only reason I know is that I tried a few years back and realized that I was asking the computer to do something that I couldn't even do. Anyone that doubts it, just listen to the NYPD feed and try to transcribe for just a minute or two.

https://www.broadcastify.com/listen/feed/32890

(edit: also, thank you for keeping this service up and running for so long, have been a regular user since the early RR days. Would love to have a comment/live chat option if your backlog is getting bare :))

>>jcims+Jb
Ok, here we go: https://feeds.talonvoice.com

Repo is here if you need to report (or just fix :D) bugs in the webapp: https://github.com/lunixbochs/feeds

>>lunixb+me
Whoa this is awesome! Love the option to fix a transcription, should hopefully help with training if you get some traction.

>>jcims+De
Thanks! I did a lot of tests and no existing ASR I found could do it to 100%, so I'm using the best ASR I could find and hoping users will help with transcriptions if they want to see it succeed and scale.