nshm

12 ms·

91.
by nshm 5y ago
Well, you can do some basic things with this but honestly RPi4 is pretty much underpowered CPU for any serious AI. You can run some optimized networks probably even 200 fps like TinyYOLO linked below but it will always be a tradeoff between
92.
by nshm 6y ago
Some repos for interested: https://github.com/FederatedAI/FATE https://github.com/bytedance/fedlearner https://github.com/adap/flower https://github.com/chaoy
93.
by nshm 6y ago
Please consider integration with Vosk https://github.com/alphacep/vosk for offline speech recognition. It should be a good fit.
94.
by nshm 6y ago
For open source offline TTS with more or less recent algorithms you can check https://github.com/TensorSpeech/TensorFlowTTS
95.
by nshm 6y ago
Its ok, but what confuse me that developers are not very responsive. I submitted a question on github some time ago on how to integrate our speech recognition services to microg https://github.com/microg/GmsCore/is
96.
by nshm 6y ago
Yeah, modern open source speech recognition like Vosk can have the cost like $2c per hour (70 times less than Google STT cost $1.4/hour) and should be just enough for search. @wenbin do you need any help with it?
97.
by nshm 6y ago
The secret of RNN-T is that it is __extremely__ hard to train, it is very unstable. On a single GPU you'll spend years to train reasonable quality model, moreover a streaming one. Streaming training requires teacher-student setup. That
98.
by nshm 6y ago
Other piece of Russian literature is also relevant: https://en.wikipedia.org/wiki/Heart_of_a_Dog The story of turning dog into a man by transferring a part of the human brain
99.
by nshm 6y ago
5 years ago https://news.ycombinator.com/item?id=9950179
100.
by nshm 6y ago
Every time i fire a data scientist the earnings grow
101.
by nshm 6y ago
Meanwhile Google trains their systems on 3 million hours of Youtube audio gaining 30% better accuracy: https://arxiv.org/abs/2010.12096 It improves much faster than you might think.
102.
by nshm 6y ago
Feels like wav2letter will not be actively developed anymore. Understandable since it is hard to compete with Pytorch with a custom NN toolkit. Any plans to move to Pytorch/Tensorflow?
103.
by nshm 6y ago
Jasper is very old and not very accurate. For offline recognition on RPi try something like Vosk https://github.com/alphacep/vosk-api
104.
by nshm 6y ago
Time to invest into open source smart devices which work fully offline. We have Vosk for offline speech recognition, there is TensorflowTTS with very good synthesis quality even on mobile phones. NLU and dialog management is also not a big
105.
by nshm 6y ago
Same in chrome on macbook, suddenly it starts to eat 200% cpu and only chrome restart helps. Battery is killed already. Did anyone notice that?
106.
by nshm 6y ago
And Microsoft announced it too https://docs.microsoft.com/en-us/azure/cognitive-services/sp...
107.
Google speech-to-text on premise in Anthos(cloudblog.withgoogle.com)
2 points by nshm 6y ago|1 comments
108.
by nshm 6y ago
GPT2 is not a good language model but there are things like XLM. Mozilla DeepSpeech doesn't support XLM rescoring, other toolkits do and it gives great improvement in accuracy. If you care about accurate transcription you'd better
109.
by nshm 6y ago
Not a big problem given so many alternatives around like E.g. some very active projects are: * Kaldi ( https://github.com/kaldi-asr/kaldi/ ) obviously, probably the most famous one, and most mature one. For standard
110.
by nshm 6y ago
Audio recording API in Unity is awful, you can only record a short clip on button press and thats it. No continuous listening, no sample rate configuration, no realtime. And processing thread changes from version to version. We had big trou
111.
by nshm 6y ago
It is going on, see https://github.com/learning-at-home/hivemind and https://github.com/bytedance/fedlearner
112.
by nshm 6y ago
Is your current data public? Do you have sufficient amount of untranscribed data (1000+ hours)? We could help you.
113.
by nshm 6y ago
It is explainable given that there was a single developer working on TTS. It is hard to compete with big academic teams/industry players this way. I also believe Mozilla team was restricted by a lack of computing resources. They had ju
114.
by nshm 6y ago
Recent more or less reasonable one is https://github.com/TensorSpeech/TensorFlowTTS , it implements all the latest algorithms. For simple business books it will be ok, for emotional fiction probably not there yet.
115.
by nshm 6y ago
There are many cloud-less Alexas already, a good one is rhasspy https://rhasspy.readthedocs.io/en/latest/ , it is not based on deepspeech though.
116.
by nshm 6y ago
Try https://github.com/alphacep/vosk-api . It supports 10 languages, works on Android and RPi and also has big and more accurate server models. Other good ones are https://github.com/daanzu/kaldi-ac
117.
by nshm 6y ago
CV is also in some not quite clear state https://discourse.mozilla.org/t/mozilla-org-wide-updates-imp...
118.
by nshm 6y ago
You have wrong premise here. First of all, current Mozilla data is almost useless for training because it is carefully read speech. You do not need much of it, even accents. If you add 1000 hours of CV data to 1000 hours of random data the
119.
by nshm 6y ago
The most important point Firefox people forget: it is better to be a good browser than bad Alexa.
120.
by nshm 6y ago
> Deepfake technology is not sophisticated enough to mimic an entire phone call with someone. With modern voice conversion technology it is perfectly possible actually.

More ›