Depending on your objective, noisy data might be useful.
I'd like LibreASR to also work in noisy environments, so training on data that is noisy should already help a bit with that.
But yeah - stammering and delays are present not only in Common Voice but also Tatoeba and YouTube.
Audio boards based on ESP32 boards are quite under the radar and have lovely features for just a few bucks. Running LibreASR on a RPi should also be feasible soon.
Data and compute are the largest hurdles. I only have one GPU and training one model takes 3+ days, so I am limited by that. Also, scraping from YouTube takes time and a lot of storage (multiple TBs).
Mozilla Common Voice data is already used for training.
I haven't trained on LibriSpeech exclusively, but yes, the perf on LibriSpeech dev is quite bad, around ~60.0 WER. If the poor alignment of yt captions is the issue, maybe concatenating multiple samples helps a bit.
I've been working on this for a while now.
While there are other on-premise solutions using older models such as DeepSpeech [0], I haven't found a
deployable project supporting multiple languages using the recent RNN-T Architecture [1].
Please note that this does not achieve SotA performance.
Also, I've only trained it on one GPU so there might be room for improvement.
Edit: Don't expect good performance :D this is still in early stage development. I am looking for contributers :)