Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Show HN: LibreASR – An On-Premises, Streaming Speech Recognition System

github.com

Post reply on HN
Print view View original

41–50 of 73 posts

Re: Show HN: LibreASR – An On-Premises, Streaming Speech Recognition System

#41

Post by whimsicalism »

fareesh wrote: post #22

Is there an open-source or paid SDK/API that I can use to create a group voice chat mobile app with "live" transcription? Or something that can plug-in to a system like this? I looked at Twilio but they seem to only offer a means to do it on their VOIP/SIP product.

> open-source or paid SDK/API that I can use to create a group voice chat mobile app with "live" transcription? Or something that can plug-in to a system like this?

Yes, Google, Amazon, Microsoft all offer streaming solutions (wouldn't recommend Amazon's however, might recommend Microsoft over Google). wav2letter from FB is the only open-source framework worth looking at, deepspeech is not a seriously usable framework.

Re: Show HN: LibreASR – An On-Premises, Streaming Speech Recognition System

#43

Post by sbr464 »

fareesh wrote: post #22

Is there an open-source or paid SDK/API that I can use to create a group voice chat mobile app with "live" transcription? Or something that can plug-in to a system like this? I looked at Twilio but they seem to only offer a means to do it on their VOIP/SIP product.

Telnyx has media forking, the ability to clone a media stream in real time without affecting the original call. It allows receiving the stream directly and operating on it without latency.

Not sure if relevant though, it's using their SIP product also. If the original service isn't using Telnyx, you could get creative and have a Telnyx shadow user join the group call to receive the stream, etc.

Re: Show HN: LibreASR – An On-Premises, Streaming Speech Recognition System

#44

Post by lunixbochs »

iceychris wrote: post #20
blackcat201 wrote: post #16

Earlier quoted context omitted.

Do you get good results when adding scraped youtube audio? My model performance on LibriSpeech dev drops a bit when adding youtube audio to the training dataset ( my guess is likely due to poor alignment from auto generated captions ).

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.

You should consider realignment; maybe start with something like DSAlign or my wav2train project.

Re: Show HN: LibreASR – An On-Premises, Streaming Speech Recognition System

#45

Post by Shared404 »

klysm wrote: post #31
jack_pp wrote: post #29

Earlier quoted context omitted.

Why does it take a lot of data? Afaik you can select lower quality in youtube-dl but you don't even need video do you?

I know you can scrape only audio from YouTube with YouTubeDL but it’s somewhat annoying

I use something akin to

    'alias downloadmusic='youtube-dl --extract-audio --audio-quality 0 --extract-metadata'
in my .bashrc

I find that helps with the annoyance of downloading things off of YT. This is for music obviously, but there's an option to download subtitles as well.

EDIT: Typed this from memory, there may be errors in the alias.

Re: Show HN: LibreASR – An On-Premises, Streaming Speech Recognition System

#46

Post by woodson »

whimsicalism wrote: post #41
fareesh wrote: post #22

Is there an open-source or paid SDK/API that I can use to create a group voice chat mobile app with "live" transcription? Or something that can plug-in to a system like this? I looked at Twilio but they seem to only offer a means to do it on their VOIP/SIP product.

> open-source or paid SDK/API that I can use to create a group voice chat mobile app with "live" transcription? Or something that can plug-in to a system like this? Yes, Google, Amazon, Microsoft all offer streaming solutions (wouldn't recommend Amazon's however, might recommend Microsoft over Google). wav2letter from FB is the only open-source framework worth looking at, deepspeech is not a seriously usable framewor…

Check out Kaldi. It's a toolkit rather than a ready-to-deploy service but has some solid pretrained models and recipes for training your own. You can use various existing projects for deployment, e.g. vosk-server (also for on-device) which comes with models for various languages and accents and has an excellent support channel via telegram. Quite frankly, despite not being "end-to-end", you'll get much much better results in practice.

Re: Show HN: LibreASR – An On-Premises, Streaming Speech Recognition System

#47

Post by whimsicalism »

woodson wrote: post #46
whimsicalism wrote: post #41

Earlier quoted context omitted.

> open-source or paid SDK/API that I can use to create a group voice chat mobile app with "live" transcription? Or something that can plug-in to a system like this? Yes, Google, Amazon, Microsoft all offer streaming solutions (wouldn't recommend Amazon's however, might recommend Microsoft over Google). wav2letter from FB is the only open-source framework worth looking at, deepspeech is not a seriously usable framewor…

Check out Kaldi. It's a toolkit rather than a ready-to-deploy service but has some solid pretrained models and recipes for training your own. You can use various existing projects for deployment, e.g. vosk-server (also for on-device) which comes with models for various languages and accents and has an excellent support channel via telegram. Quite frankly, despite not being "end-to-end", you'll get much much better re…

I collected custom audio and had it transcribed by hand for cash, then evaluated it on wav2letter and vosk. At least for that domain, wav2letter outperforms vosk.

Re: Show HN: LibreASR – An On-Premises, Streaming Speech Recognition System

#48

Post by woodson »

whimsicalism wrote: post #47
woodson wrote: post #46

Earlier quoted context omitted.

Check out Kaldi. It's a toolkit rather than a ready-to-deploy service but has some solid pretrained models and recipes for training your own. You can use various existing projects for deployment, e.g. vosk-server (also for on-device) which comes with models for various languages and accents and has an excellent support channel via telegram. Quite frankly, despite not being "end-to-end", you'll get much much better re…

I collected custom audio and had it transcribed by hand for cash, then evaluated it on wav2letter and vosk. At least for that domain, wav2letter outperforms vosk.

Good for you, it's the only way to know which tool works best in your case. I did the same for my use case and arrived at the opposite conclusion.

What most people don't realize is that it heavily depends on your use case and domain whether any given model/algorithm will work better.

Re: Show HN: LibreASR – An On-Premises, Streaming Speech Recognition System

#49

Post by lunixbochs »

Can you post WER per dataset? Bucketing all of the WER together means you can only directly compare to models that are validated on the exact same combination datasets. This excludes all other ASR systems from comparison, as well as your own models if you decide to add validation data in the future.

Re: Show HN: LibreASR – An On-Premises, Streaming Speech Recognition System

#50

Post by woodson »

iceychris wrote: post #4

Hey HN! 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 t…

You can also check out https://github.com/TensorSpeech/TensorFlowASR for inspiration (not my project, not involved). It implements streaming transformers and conformer RNN-T (but in TF2). Deployment on device as TFLite. So far, there aren't many usable pretrained models available (just LibriSpeech), but with some work it could turn out quite nicely.
Post reply on HN