woodson

12 ms·

151.
by woodson 3y ago
Not sure, but the mouse was invented at SRI.
152.
by woodson 3y ago
If you’re interested in ML training, to save you some time, RTX 4080 Super and RTX 4070 Ti Super have 16GB GDDR6X, the RTX 4070 Super has 12GB.
153.
by woodson 3y ago
The question is whether to legislate AI or spreading of brazen falsehoods.
154.
by woodson 3y ago
That’s true, one has to design the model for the deployment target. Especially avoiding in-place tensor operations and python control flow helps for tracing.
155.
by woodson 3y ago
I’m somewhat concerned that it doesn’t support ANE from the get-go (they list CPU and GPU, for now). Not even for inference only. Other frameworks have (non-python) solutions for mobile in place, e.g. tflite, libtorch, and onnxruntime.
156.
by woodson 3y ago
The article is very thin on actual technical details. It doesn’t describe how the spectrograms are processed or details of the neural network they’re using (other than using convolutions, which is pretty standard in CV/audio/speec
157.
by woodson 3y ago
It would be useful to be able to change the spectrogram parameters to achieve what you’re suggesting, but it’s likely that this might cause the model to misbehave if it was trained only on spectrograms of audio samples that weren’t frequenc
158.
by woodson 3y ago
I remember when they (Mozilla’s CV team) solicited feedback before they got started, I brought up that issue and proposed a different approach to gathering conversational speech data, but it wasn’t picked up. The belief that it’s better to
159.
by woodson 3y ago
I’d go with a remote dev solution. Training/finetuning of large models requires much more resources anyway, so the GPUs in the local machine would be unused most of the time.
160.
by woodson 3y ago
A simple correlation of audio chunks from microphone and from the TTS should be enough to tell which parts in the input stream are re-recorded TTS. Much simpler, no?
161.
by woodson 3y ago
Cultural factors are one thing, the other is just that there are only ~5.8 million native Finnish speakers, so a lot of books, movies, music, etc. are not available in the language for purely economic reasons. There are over 90 million Germ
162.
by woodson 3y ago
You’re saying there is no equivalent to Ladenzeitgesetz in Germany?
163.
by woodson 3y ago
Very nice library. I used it some time ago to implement a module that applied ASR and other audio analytics to audio streams in real time and relayed extracted information via SIP messages. As with many projects in that space, it doesn’t co
164.
by woodson 3y ago
Perhaps it’s charging client A for work performed (i.e., time spent) on client B.
165.
by woodson 3y ago
It was primarily being used to train TTS models (see https://github.com/neonbjb/tortoise-tts ), which largely fit into a single GPUs memory. So, for data parallelism, x8 PCIe isn't that much of a concern.
166.
by woodson 3y ago
It's roughly on par with NVIDIA RTX A6000, which also has 48GB VRAM.
167.
by woodson 3y ago
It wasn’t the donations. I bought a GTX 580 back then to train DNNs for speech recognition. There simply were no other reasonable options around.
168.
by woodson 3y ago
But that’s not one-party consent. The ring cameras aren’t a party.
169.
by woodson 3y ago
As someone not from the US, it’s not a surprise at all. But if someone says a foreign company is making their employees work on a holiday weekend, I don’t fault them for not explaining what specific holiday it is.
170.
by woodson 3y ago
I generally agree with your sentiment, but Twitter is a US company with most of their employees residing in the US (at least presumably?). So in this case it’s reasonable to assume they meant it’s a holiday in the US.
171.
by woodson 3y ago
Google open-sourced (Apache 2.0) the Switch Transformers C-2048 model (1.6T parameters for 3.1 TB): https://huggingface.co/google/switch-c-2048
172.
by woodson 3y ago
But CUDA kernels are everywhere, not just in high-level frameworks. Look at DeepSpeed, for example, which is used in training LLMs.
173.
by woodson 3y ago
EnCodec has since been relicensed to MIT.
174.
by woodson 3y ago
fairseq has existed for many years and is Fb’s (now Meta’s) AI research group’s sequence-to-sequence modeling toolkit. It’s not specific to this MMS model.
175.
by woodson 3y ago
It’s mostly there in https://github.com/lucidrains/audiolm-pytorch#hierarchical-t... . They just used FAIRs EnCodec ( https://github.com/facebookresearch/encodec ) instead of soundstream.
176.
by woodson 3y ago
Just to add a datapoint: the main audioLM based models (not the BERT embedding part) fully utilize an RTX 2080 Ti.
177.
by woodson 3y ago
Have you considered K2/Sherpa for ASR instead of ctranslate2/faster-whisper? It’s much better suited for streaming ASR (whisper transcribes 30 sec chunks, no streaming). They’re also working on adding context biasing using Aho-Cor
178.
by woodson 3y ago
I hear the Moselle river valley is nice for bike tours, though not quite as long.
179.
by woodson 3y ago
Of course, but going from DARPA PAL and CALO to Siri is like going from Mikolov et al. (2013)’s word2vec to ChatGPT.
180.
by woodson 3y ago
Sorry to be pedantic, but Siri was released 2011 and Google Home 2016. With all the recent progress on LLMs, people forget that these “old” assistants are rather recent developments.

More ›