nshm

10 ms·

61.
by nshm 4y ago
I think they didn't use WSJ for training, only for evaluation. Paper includes WSJ under "Evaluation datasets"
62.
by nshm 4y ago
And tensorflow
63.
by nshm 4y ago
First of all thank you for your nice research! It is really inspiring. > Also in Table 6, you see that Facebook's wav2vec 2.0 XLS-R went from 12.06% without LM to 4.38% with 5-gram LM. It is probably Jonatas Grosman's model, no
64.
by nshm 4y ago
The model knows the recognition text well and demonstrates good results because of it. If you test the same model on some unrelated speech which model didn't see yet the results will not be that great. Error rate might be significantly
65.
by nshm 4y ago
This model seems strongly overtrained on CV test set. Usually improvement from LM rescoring is just 10% relative. In the paper https://arxiv.org/pdf/2206.12693.pdf the improvement is from 10.1% WER to 3.64% WER (Table
66.
by nshm 4y ago
Amazing given ffi-napi is broken for several years already https://github.com/node-ffi-napi/ref-napi/issues/54
67.
by nshm 4y ago
Automatic speech recognition for search and redaction too.
68.
by nshm 4y ago
I don't know of easy to use software for this, but algorithms are there like in https://benedikt-bitterli.me/astro/
69.
by nshm 4y ago
We all know Bay Area is a great place to live geographically ;) Try to find black zone around Chicago for example.
70.
by nshm 4y ago
Interesting this field is changed significantly last 20 years. One thing is light pollution need to buy a big telescope if you can't transport it to the dark place. You have to drive 100+ from a big city. Second is advanced image proce
71.
by nshm 4y ago
For many domains active learning is not that efficient actually. The promise is that you make a subset of labels and train on them the model with the same accuracy. The reality is that in order to estimate long tail properly you need all th
72.
by nshm 4y ago
Commonvoice is actually not very good test set. Their texts are very specific (mostly wikipedia and such) and also the texts overlap between train and test which leads to overtraining of most transformer models. If you test on variety of do
73.
by nshm 4y ago
Not when you are on Macbook. Don't suggest external drives please ;)
74.
by nshm 4y ago
For me SDK sizes are showstopper. Mingw packages are probably 100-200 Mb total. To build anything native one has to install Windows (20+Gb in modern version and then Visual Studio 40+Gb). Not so easy to fit it all on an SSD drive. Xcode als
75.
by nshm 4y ago
The standard operation for FSM to compose knowledge sources is composition. You can compose different level graphs and differentiate them too. Awni's paper talks about that https://openreview.net/pdf?id=MpStQoD73Mj . Th
76.
by nshm 4y ago
Seems like in this differentiable FSM implementation you only differentiate edge weighs, you can not introduce or change nodes. So it is quite far from differentiation of programs.
77.
by nshm 4y ago
Did you try Vosk https://github.com/alphacep/vosk-api or Wenet https://github.com/wenet-e2e/wenet
78.
by nshm 5y ago
It is clear Google has issues with indexing Github. Couple years ago they somehow retrieved github mobile page instead of desktop page. As a result, no comments, no readmes, no code indexed at all. A solution is to move most of the document
79.
by nshm 5y ago
For offline transcription try Vosk https://github.com/alphacep/vosk-api
80.
by nshm 5y ago
In general, language detection is surprisingly hard. There is LSTM-based implementation https://github.com/AU-DIS/LSTM_langid which should be better than ngrams.
81.
by nshm 5y ago
Speech input with something like Vosk is a great idea too.
82.
by nshm 5y ago
We are working on punctuation/capitalization. Models are ready but not yet integrated. German BERT-based model is available at https://alphacephei.com/vosk/models#punctuation-models for example. Should be ready so
83.
by nshm 5y ago
Exactly. Moonman T5 is a bomb, highly recommended. Many other manufacturers are also very good. Ignore people trying to convince you a pen should cost $400.
84.
by nshm 5y ago
Any of these with voice control? Preferably without internet connection.
85.
by nshm 5y ago
The simplest form of unsupervised learning or self-teaching is clustering. And yes, it works without labels.
86.
by nshm 5y ago
It is true when your task if relatively simple and you can create labels to cover most of the data cases. For long tail problems like vision or NLP you can't label everything and semi-supervised learning helps a lot.
87.
by nshm 5y ago
Huggingface ASR models are not really recommended. The simple fact they don't use beam decoder with LM makes them much less accurate for practical applications. If you compare them to setups like Nemo + pyctcdecode, they will be 30% le
88.
by nshm 5y ago
Data cost plunges these days with self-supervised and semi-supervised learning. You don't need annotated and clean data anymore, there is abundance of it. Projects like Voxpopuli or Gigaspeech with 400 thousand hours (100 times more th
89.
by nshm 5y ago
Did you try Vosk Indian English model? It is specifically built for Indian accent English https://alphacephei.com/vosk/models/vosk-model-en-in-0.4.zip In case you want more accuracy you can share a file with an ex
90.
by nshm 5y ago
Thank you. I deeply appreciate you mention our efforts. We spend quite some time and knowledge to build accurate speech recognition. Not that easy to get as much mentions as Mozilla, so we are thankful for every single one!

More ›