Has anyone tried to automate transcription of the autofiles that Trunk-Recorder creates? I was thinking that if we could make this work, you could possibly then set alerts for keywords, etc. Thoughts? Is there something else that can do this already?
Trunk-Recorder Transcription?
- Thread starter dbuttry
- Start date
-
- Tags
- trunk-recorder
- Status
Just started looking at that. Looks like Trunk-Recorder outputs a .wav in 8000hz mono, and 16000hz would be a much better sample rate for some kind of speech/dictation software. Poking at the code a bit, I can't figure out how to make that change. Not sure how the digital recorder works.
Specifically this: /trunk-recorder/p25_recorder_decode.cc
wav_sink = gr::blocks::nonstop_wavfile_sink_impl::make(1, 8000, 16, true);
It would appear it's much more complex than just this line (double speed audio output if you change it, so there must be other hacks required). I'm bad at C, and I haven't figured out how the end .wav gets dumped out. Suggestions? Am I looking at the wrong thing?
So far DeepSpeech (linux) looks a bit more promising and active than Julius (linux) for the wav-to-text thing. I'd imagine you'll be working on the "model" a bunch trying to train it to make it work. Multiple voices, not very clear.....going to be an issue.
Anyone else try this?
Specifically this: /trunk-recorder/p25_recorder_decode.cc
wav_sink = gr::blocks::nonstop_wavfile_sink_impl::make(1, 8000, 16, true);
It would appear it's much more complex than just this line (double speed audio output if you change it, so there must be other hacks required). I'm bad at C, and I haven't figured out how the end .wav gets dumped out. Suggestions? Am I looking at the wrong thing?
So far DeepSpeech (linux) looks a bit more promising and active than Julius (linux) for the wav-to-text thing. I'd imagine you'll be working on the "model" a bunch trying to train it to make it work. Multiple voices, not very clear.....going to be an issue.
Anyone else try this?
After a little more time ---> The 8000 Hz recording is just how the P25 is voice decoded into the wav file and just how it works, but the real battle is the modeling, audio cleanup/filtering, and speech recognition.
They speak fast on the radio, the audio isn't normalized, and not very clear in a lot of calls. The keywords that are weighted -- that a default model looks for -- isn't exactly normal speech patterns, so I get some really awesome and funny results on DeepSpeech. It's trying. It just doesn't know about sentence fragments and radio "banter".
It is going to be a keyword, machine learning, and model training exercise, so a machine with a lot of horse power, with a decent graphics card you can leverage the GPU's on is going to be required to make a model. Nvidia looks to be the most supported for this. Maybe it's possible to get close(r).
They speak fast on the radio, the audio isn't normalized, and not very clear in a lot of calls. The keywords that are weighted -- that a default model looks for -- isn't exactly normal speech patterns, so I get some really awesome and funny results on DeepSpeech. It's trying. It just doesn't know about sentence fragments and radio "banter".
It is going to be a keyword, machine learning, and model training exercise, so a machine with a lot of horse power, with a decent graphics card you can leverage the GPU's on is going to be required to make a model. Nvidia looks to be the most supported for this. Maybe it's possible to get close(r).
polkaroo
Missed him again!
- Joined
- Apr 14, 2002
- Messages
- 408
There's also Microsoft Azure Speech to Text – Audio to Text Translation | Microsoft Azure and Google Speech Speech-to-Text: Automatic Speech Recognition | Google Cloud though I'm sure (obscure) street names will give some interesting results.
I wanted to do this, as well, as a method for filtering down the calls I choose to listen to from my system. I recall trying DeepSpeech, and not getting great results from it — and, IIRC, making adjustments meant retraining a new model, which I really don't know enough about, have enough annotated training data for, or have patience for 
I tried a bunch of other things (that I've forgotten now — I did this Dec 2020) with mixed results. I found Google's Speech-to-Text API to give the best results. It also offers the ability to make adjustments (i.e. adding and/or weighting keywords) and provides position, duration, and confidence for each word it transcribes. I had the best luck using the phone call model. Enhanced mode had mixed results. But even with all my tinkering, it was never great.
Here's an example from my notes:
For reference, I've attached the script I was playing with — it needs Python 3.10 ('cause I use generic typing on builtin collections), and was developed with dependencies
I tried a bunch of other things (that I've forgotten now — I did this Dec 2020) with mixed results. I found Google's Speech-to-Text API to give the best results. It also offers the ability to make adjustments (i.e. adding and/or weighting keywords) and provides position, duration, and confidence for each word it transcribes. I had the best luck using the phone call model. Enhanced mode had mixed results. But even with all my tinkering, it was never great.
Here's an example from my notes:
For reference, I've attached the script I was playing with — it needs Python 3.10 ('cause I use generic typing on builtin collections), and was developed with dependencies
colorama==0.4.4 and google-cloud-speech==2.0.1- Status
