woodson

13 ms·

121.
by woodson 2y ago
Yes, vLLM does (though marked experimental): https://docs.vllm.ai/en/latest/models/vlm.html
122.
by woodson 2y ago
Portland embraces it, see for example the Ned Flanders Crossing, on Flanders St, no less: https://en.m.wikipedia.org/wiki/Ned_Flanders_Crossing (The Simpsons character is named after the street.)
123.
by woodson 2y ago
The concern may be about what effect this will have on future papers (just like news headlines engineered for clickbait).
124.
by woodson 2y ago
Not OP, but I agree that this could lead to questionable learning outcomes, especially since Whisper isn’t that good for low-resource languages. It’s probably fine for languages like English/Spanish/Mandarin, though.
125.
by woodson 2y ago
What I found is that, for cross-language use-cases, this often just applies the intonation of the “context” sample to the created sample, which, if they are from different languages, usually gives the wrong result (in the sense that it soun
126.
by woodson 2y ago
And like with most other competition/benchmark, the result is likely optimizing for the benchmark and not the wider goal ;-). It’s difficult to get a serious effort without people trying to game the benchmark.
127.
by woodson 2y ago
It must have been a different program, as this one hasn’t started yet, but perhaps by another program by the same program manager.
128.
by woodson 2y ago
It’s a shame that the article didn’t mention AWQ 4-bit quantization, which is quite widely supported in libraries and deployment tools (e.g. vLLM).
129.
by woodson 2y ago
You can run multi-node with tensor parallel plus pipeline parallel inference, e.g. with vLLM ( https://docs.vllm.ai/en/latest/serving/distributed_serving.h... ).
130.
by woodson 2y ago
I wonder if AutoAWQ works out of the box, given no architectural changes (?). That would be most straightforward together with vLLM for serving.
131.
by woodson 2y ago
Thanks, I haven’t come across cppyy! But I’ve worked with pybind11, which works well, too.
132.
by woodson 2y ago
Ah my bad, you’re right. I think it doesn’t do V2V directly, but can use reference audio to guide TTS.
133.
by woodson 2y ago
Try StyleTTS2. You will still have to experiment with the settings a little to get the right level of adherence to the reference speaker’s voice and the emotion content.
134.
by woodson 2y ago
That reminds me of the findings of Google’s paper on EfficientT5 ( https://arxiv.org/abs/2109.10686 ). They refer to it as “DeepNarrow”.
135.
by woodson 2y ago
ASR systems already use language models during decoding, though mostly not large decoder-only LLMs. However, incorporating LLMs into ASR is currently at the center of a lot of research, e.g. using a speech encoder like wav2vec 2.0 or the wh
136.
by woodson 2y ago
Of course, all I said was that the table doesn’t account for it. Aside from that, judge for yourself: https://en.m.wikipedia.org/wiki/Colombian_Spanish
137.
by woodson 2y ago
These differences go back much further in the past, so FARC has nothing to do with it in Colombia (it’s spelled with an “o”). There’s a large linguistic diversity within these countries, which that table doesn’t reflect or account for.
138.
by woodson 2y ago
A lot of words and expressions in the old Viennese dialect are Yiddish, though that dialect has all but disappeared from day to day interactions (you can find it in old movies and recordings).
139.
by woodson 2y ago
Yes, it should work just fine with mokutils.
140.
by woodson 2y ago
The human driver is liable, the machine is not (or not in the same sense).
141.
by woodson 2y ago
Well, strictly speaking modern LLMs are transformers (not counting state space models like Mamba, for the moment). Not sure what hidden Markov chains have to do with modern LLMs.
142.
by woodson 2y ago
The wheels are cached (check ~/.cache/pip/wheels on Linux/Macos)
143.
by woodson 2y ago
The pretrained models aren’t MPL licensed, though.
144.
by woodson 2y ago
Query expansion has been done since forever, long before even word2vec and other semantic embedding models, e.g. using WordNet (not a DNN, despite its name), see https://lucene.apache.org/core/3_3_0/api/contri
145.
by woodson 2y ago
Replacing the codebook approach with a statistical/DNN is more likely to give higher accuracy than getting rid of mfccs as spectral representation (at least in general ASR). (Arguably, using Mel spectra was the least controversial desi
146.
by woodson 2y ago
The first few editions of the “Nuke” music festival were also held there in the early aughts: https://de.m.wikipedia.org/wiki/Nuke_(Musikfestival)
147.
by woodson 2y ago
Especially for small models I had very bad results for use in translation. Even trying all kinds of tricks didn’t help (apparently prompting in the target language helps for some). Encoder-decoder models such as FLAN-T5 or MADLAD-400 seemed
148.
by woodson 2y ago
Since EPUBs are just zipped HTML files, I guess rendering can be done more efficiently by the browser itself rather than by custom rendering on an HTML canvas.
149.
by woodson 3y ago
Traditional codecs also use a lot of “magic” tables with numbers (see e.g. AMR codecs used in GSM telephony). I think this codec could be optimized to run relatively efficiently on the various AI accelerator chips modern phones have, which
150.
by woodson 3y ago
The difference is that Codec2 encodes audio at 8 kHz sampling rate while this TSAC codec encodes audio at 44.1 kHz, which makes a pretty big difference in terms of audio fidelity.

More ›