Transcribing audio from scanner into text.

[Arcreef]

Hi all I'm looking into coding a script that will automatically put the audio over the airways into text files. You'd be able to go back and read all the transmissions made. I like to record what goes on in my area then listen to it at night. Having it in text would make it easier to scan through and I'm also hard of hearing so I wouldn't have to play it back 5 times to figure out what they said.

Right now I can take the audio files from my SDS100 and drag them into a transcription program and it will turn it into txt. But what I want is to leave my scanner hooked to the computer, once the SDS100 or SDS200 record 100 transmissions, it creates folder for those 100 recording. The script will scan a folder for any new files, once it detects a new folder it will send those 100 audio files to the transcription program, which will convert the audio into text. Then take the text and automatically enter it into an excel file and add a time stamp to each text and will highlight the text box in red if a tone out accompanied the transmission. Then it will save the file in a drop box folder or 1drive so I can access the transcript text anywhere I'm at on my phone.

Has anyone ever heard of a program that would do this? I don't want to spend time on it if someone has already done it before.

Open to any added suggestions or critique.

Thanks!

 

[jtwalker]

@ProScan does the recording and tagging with timestamp part. Not transcription part though, nor moving to your DropBox.

 

[fxdscon]

Open to any added suggestions or critique.

Might be some help in this recent thread here:

Audio file to text?

Is there a proscan add on or other software you can add to convert audio files to text to stream host on web server? Thanks in advance.

forums.radioreference.com

 

[wtp]

if you can't make it out...how is a program supposed to ?
i have seen some live news broadcasts that really had a problem with 'accents' so to speak.
recognize speech
wreck a nice beach
comes to mind and i can't think of where i got that.

 

[jasonk]

What System are you monitoring ?

 

[Arcreef]

if you can't make it out...how is a program supposed to ?
i have seen some live news broadcasts that really had a problem with 'accents' so to speak.
recognize speech
wreck a nice beach
comes to mind and i can't think of where i got that.

Good point I'll look into which transcription software had the best training. I'd assume it'll still do better than a half deaf guy. I'll have to experiment first though to see 8f that's an issue

 

[Arcreef]

What System are you monitoring ?

P25 simulcast, mostly miami police southern Station. I know it's not perfect, whichnis another reason I wanted to turn it into text, bit maybe it will have trouble making out words also. I don't need it perfect though.

 

[fog]

I am interested in this topic. I experimented briefly and had very bad results. I was trying cloud services; AWS Transcribe and MS Azure's Speech to Text both struggled to give me results that weren't largely gibberish.

I had the idea to write something that would feed it streams of several local scanner feeds, partially for easier search, but largely for keyword alerting: I don't necessarily want to listen to the fire department chasing false alarms all day, but alert me if "working fire" or "second alarm" gets said.

You might do better if you can have it run on each radio transmission individually. One of the problems I had was that, given a recording of 30-60 minutes, it tended to just mash it together, maybe erratically giving paragraph breaks, but absolutely never formatting it as a transcript/dialog with multiple speakers.

I also think it really struggles with P25 or other digital voice formats, and has probably been trained on phone calls as the "worst" quality.

I just went digging through an email I sent a friend talking about my experience; here is an actual example of a 30-minute Broadcastify recording of my local police department (P25 conventional), as transcribed by AWS Transcribe:

Okay, so she's a guy. Cattle from Got it. Total seven. So Okay, I did call him back, and he said he's not only he's all set, and thank you. Okay. Yeah. Seven. Oh, I'm going to have over 17. Okay, that was 204 university of that truck and was at her apartment number. And I also apologize. I don't have a computer. No apartment. Yeah, he's going to be out there. All right. I want Well, your current car one. You have a UK on animal control. About 30 minutes. The girl. Enough. Hello? Go ahead. I'll be checking on a couple at 1. 55 minutes. 15. Okay. You all set? There a problem? Go home. I'm just leaving the ark and I'll head over there something for but all right, class, she's going to ST voluntarily. Okay. Eating calories and moved along. I'll be clear where I thought it was hard to. Yeah. 80 Hampshire Street is going to be the Yellow House. Loud music. All right. For 11. Hello. Okay. Okay. To 80 Hamshahri tree for loud music before so forth. 18 Morton Street for allow music. It's going to be in the back yard. Okay, hold on. 19. 19. Small harsh disregard for the car. 10. 10 clear In our can five worked for ST. It's going to be a harassment. Blackmail was T shirt blue. Jean has about 55 He's outside harassing Carlo. So world a car. Five terrified Western street. Yes, please. 22. I'll take that with Okay, five. Disregard profile copies. Thank you to 11. You can cancel. Durable music wasn't unreasonable, but I turned it down. Well, they won't. They're supposed help from it. Well, number Summer South Street that summer. It's all that drama as three is going off at Westminster. Thank you. Animal control. The little mhm. You please sign in? Yes. Really? Mhm. You can clear whether you can have mhm Angel. Very Georgia visual. Excel literacy. California to Costa. Good. We'll watch for that For somebody. College residual Glad. Can I get trinity to South Common? Right next to the pool or sharps removal. Mhm problem. Proposal, Cowboy. Tomorrow morning, far as well. You can show me clear from the back of the election. First

Some of it I can kind of make sense of. "You have a UK on animal control" was surely, "Do you have an ETA on Animal Control?", for example. (I think the apology about not having a computer was likely accurate, and was an officer whose MDT was down.) And I think "blackmail" was a description of a person's race and gender.

I didn't save the audio file this was from, but I can assure you that, "Okay, so she's a guy," "I'm just leaving the ark," "eating calories," "The little mhm," "Very Georgia visual," and "Excel literacy" were never said. We also have no cattle or cowboys in my city, nor a Terrified Western Street.

I would be thrilled if you had better results, though. I think there's a bunch of interesting uses if you can get it to be even just generally kind of correct.

 

[IC-R20]

@ProScan does the recording and tagging with timestamp part. Not transcription part though, nor moving to your DropBox.

Free scan does it too. I have my old Dell e6400 connected to the BCT15x and logging the 'Special' itinerant service search to aid my commercial scanning quest. No idea why Uniden decided to remove those on the x36HP series, when I first got the 436 that was one of my big plans for it to do that self contained. There is the itinerant list from the database I have on my favorites but the Special search did a lot of the commercial range outside those frequencies too so I was able to get a lot more businesses.

 

[IC-R20]

Good point I'll look into which transcription software had the best training. I'd assume it'll still do better than a half deaf guy. I'll have to experiment first though to see 8f that's an issue

The dragon software isn't too bad. I used Dragon Professional to make video subtitles before and though not intentional it was also able to make out most of the words in a scene coming out of a shortwave radio with lots of static and fade.

Otherwise you might just wanna get Incident Page Network if that's in your area just be able to peruse the Incident Database for calls without having to go through lots of chatter in-between.

 

[jasonk]

P25 simulcast, mostly miami police southern Station. I know it's not perfect, whichnis another reason I wanted to turn it into text, bit maybe it will have trouble making out words also. I don't need it perfect though.

City of Miami Florida - p25 System ? This is a Harris P25 system

 

[Ubbe]

You can't access files on the SD card while the scanner are doing scan as it has exclusive access to the file system. You will have to use an external program to log and record audio files, something like ProScan. Then you can access the files already stored on your harddrive by a program you write to fetch the files and export them as they become available and get them transcribed and save the text together with the system and TG info, that perhaps already are part of the file name of the recorded file. Maybe it can be done in something simple like Basic programming language.

If it's just one system and there's no need to have a TG info then just route the scanner audio into a transcribe program that produce a text file and that program hopefully ends a file when there's a long enough silence or perhaps each hour, and you make a script that exports those textfiles to Onedrive and delete them from their original place each hour.

But I think that Onedrive needs a lot of stuff being handled in a graphical user interface, so perhaps easier to send as an email to yourself.

/Ubbe

 

[bob550]

If you've ever tried your speech to text feature on your cellphone, you realize it's limitations especially when trying to understand unusual and unique words. And that's under relatively ideal conditions. Added to that, radio transmission quality varies considerably among different systems and agencies, with heavy dependence on the person transmitting. There have been discussions here regarding this very issue. That said, perhaps there's a professional level solution available that is in use by Public Safety agencies to transcribe their transmissions. But I would expect that to be at a considerable expense.

 

[Ubbe]

I was impressed by Microsoft's transcribe function that worked with bad scanner audio that I could hardly hear what was said. And even more so as when it had finished printing a sentence it didn't make any sense and it went back and changed the first word to a similar sounding one but with a different meaning and then the whole sentence where correct. Its AI seemed to be very powerful.

I'm also impressed by that little $30 Google Nest ball that I can have it playing loud music and it easily picks up my "Hey Google Stop" command among the music playing. When I have it in my upstairs study room I can shout "Hey Google add milk to my shopping list" when standing by the fridge in the kitchen, and when I check my phone it has added milk to my list.

/Ubbe

 

[KC1ABL]

I actually started experimenting with this as well, I use ProScan with my Undien scanner and automatically pull the new recordings from my server and upload them to a "Whishper" API that I also have running on my home network:

Whishper

Transcribe, Subtitle, Translate

whishper.net

Whishper is a opensource version of OpenAI's Whisper LLM (related to ChatGPT) and it's fully local so it doesn't require any connection to the internet to function (and it's free!). I had only minimal success but I'm still working on it.

Edit: the success I mention was mostly around the model not making very accurate transriptions like the other posters above mentioned.

 

[kb9mwr]

I've been using two tone detect to capture the first few seconds after the page tone for my community on our county system, from there that recording is submitted to google's api

https://cloud.google.com/speech-to-text/