My plan was to collect user transcription corrections on my site then train my own inexpensive models on them. The open-source speech tech I work on can do passable transcription at close to 100x faster than realtime on a quad core desktop CPU (or 200 simultaneous streams per 4-core box at 50% activity). With higher quality transcription it's closer to 10-20x faster than realtime.
For your case you could also try to push some of the computation down to the uploading machine. These models can run on a raspberry pi.
I think the biggest work for a new effort here is going to be building local language models and collecting transcribed audio to train on. However, there have been a couple of incredible advances in the last year for semi-supervised speech recognition learning, where we can probably leverage your 1 year backlog as "unsupervised training data" while only having a small portion of it properly transcribed.
The current state-of-the-art paper uses around 100 hours of transcribed audio and 60,000 hours of unlabeled audio, and I bet you could push the 100h requirement down with a good language model and mixing in existing training data from non-radio sources.
https://wiki.radioreference.com/index.php/Broadcastify-Calls
I wonder if one could mix in openstreetmap data for a location to help pick up local references. (Eventually would be cool to round trip it with a little ping when addresses/businesses are referenced).
Any thoughts on adding the ability to comment/transcribe/etc?
Your prototype is amazing! The quality of transcription is definitely better than ours via Google.
After we did some legal research we wanted to avoid storing the recordings and rather solely transcription text. Giving access to a platform for humans to verify the transcriptions and in turn train the model is a great idea.
I have started working on getting some pre-trained models set up. I am trying to implement them with wav2letter, deepspeech, kaldi, vosk, etc. - I just need to be pointed in the right direction.
Raspberry Pi's were something I was considering as well - small energy footprint and powerful enough to run these models.
Do you have any advice on ML or acoustic models to avoid? I am working with the 100 hour dataset now.
Thanks!
I have 30ish streams and keep 6 days worth, I could keep longer if you'd like to work together on this. I reached out to some of the people above, the Broadcastify guy for example, and they are, as mentioned, ready doing their own thing so didn't really care about what I wanted to share.