woodson

17 ms·

181.
by woodson 3y ago
It’s ML when you’re doing research and AI when you ask for funding ;).
182.
by woodson 3y ago
Take a look at AudioLDM ( https://github.com/haoheliu/AudioLDM ), it might be more what you expected: - Text-to-Audio Generation: Generate audio given text input. - Audio-to-Audio Generation: Given an audio, generate ano
183.
by woodson 4y ago
No, it’s not similar (other than superficially, as both use diffusion methods in some way). It uses diffusion to generate latent vectors that are entered in a neural audio codec model to produce speech.
184.
by woodson 4y ago
That's likely because it wasn't published as a paper anywhere (not even arxiv) and the author then joined OpenAI and development ceased.
185.
by woodson 4y ago
Even big ones like REI are closing their last location in Portland: https://www.oregonlive.com/business/2023/04/rei-to-close-its...
186.
by woodson 4y ago
There are some projects, such as this one: https://github.com/idiap/w2v2-air-traffic They transcribed 5000 hours of ATC speech data (available for licensing, not free).
187.
by woodson 4y ago
Back in the day, having taken some kind of statistical signal processing course would have been common before getting into neural networks. That would likely have covered a lot of intuitions.
188.
by woodson 4y ago
Well, in general, there seems to be a correlation between what news coverage focuses on and what is considered worrisome at any given time. In times of high inflation and increased fuel prices, climate change action takes a back seat.
189.
by woodson 4y ago
Whisper doesn’t do streaming transcription or translation. It’s one of its architectural disadvantages over more commonly used ASR approaches, such as transducers (see e.g. Microsoft Research’s system trained on ~350k hours that does stream
190.
by woodson 4y ago
People refer to Portland as PDX a lot. Perhaps it’s a trend.
191.
by woodson 4y ago
You probably should care if you integrate the inference code in other software packages that aren’t GPL licensed.
192.
by woodson 4y ago
The fact that AMR uses a very complex analysis-synthesis loop (with subtle differences between its 8 bit rate modes) makes speaker recognition much harder (and therefore more error prone), but perhaps they decided that they can live with
193.
by woodson 4y ago
Not GP, but what NVidia Triton can do includes - Dynamic batching while limiting latency to a set threshold - Running multiple instances of a model, effectively load-balancing inference requests. - Loading/unloading/running multip
194.
by woodson 4y ago
I suspect (though I may be wrong) that GPT-4 wouldn't provide more insight than what's in the text it was trained on. If prompted by examples of regular sound change or reconstruction of proto-indo-european (PIE) roots, it may be
195.
by woodson 4y ago
Anonymization doesn’t just mean “leave their names out”. An entire patient's medical history is in itself personal identifiable information. Instead of googling for “headache”, they now have stored a copy of every medical detail in you
196.
by woodson 4y ago
An interesting tidbit (hinted at in the article) is that there’s an active effort to reuse old (no longer used) words for new meanings. An example is sími (telephone), an old word for cord/rope. I wonder how well GPT-4 copes with the i
197.
by woodson 4y ago
It quantizes 32-bit floating point weights to 4-bit integers so that models fit into memory and run reasonably fast on a CPU by using the processor’s vector extensions.
198.
by woodson 4y ago
Well, tensor cores support int8 instructions (at least from Turing onwards), so the hardware is being used, if that’s your concern.
199.
by woodson 4y ago
One downside of pgvector is that it currently only supports one type of index (ivfflat), while others (FAISS, Milvus, qdrant, etc.) support other types of indices that can be advantageous depending on your workload (properties of vectors, s
200.
by woodson 4y ago
Part of the problem is that the confidence for “cardboard box” was probably quite high. It’s hard to properly calibrate confidence (speaking from experience, speech recognition is often confidently wrong).
201.
by woodson 4y ago
I used this in the past for a relatively complex US/foreign situation: https://www.myexpattaxes.com/ No affiliation or anything.
202.
Recursive Games with ChatGPT(gist.github.com)
1 points by woodson 4y ago|0 comments
203.
by woodson 4y ago
One point to note is that they appear to be using (pseudo-)phoneme sequences as inputs instead of characters/text, so you need a frontend that does grapheme to phoneme (G2P) conversion. I found that interesting as many previous models
204.
Recursive Games with ChatGPT(gist.github.com)
1 points by woodson 4y ago|0 comments
205.
by woodson 4y ago
Ah thanks! I looked at the page on a phone screen, where only the text and the first audio playback button are visible. My bad..
206.
by woodson 4y ago
Is it just me, or does the spoken content not correspond with the written prompt in many cases? Though, I’m sure it’s just a problem with matching the right file with the text in the HTML and not a TTS problem. [Edit: My bad, I looked at th
207.
by woodson 4y ago
I think quantization (e.g. 4-bit, https://arxiv.org/abs/2212.09720 ) and sparsity (e.g. SparseGPT, https://arxiv.org/abs/2301.00774 ) will bring down inference cost. Edit: This isn’t handwaving btw,
208.
by woodson 4y ago
Looking at the repo, it appears to be based on a model trained using K2/icefall ( https://github.com/k2-fsa/icefall ; K2 is next generation of Kaldi). It claims to be based on a LSTM Transducer model trained on the
209.
by woodson 4y ago
Looks like this isn't based on OpenAI's whisper, but on a model trained with RNNT loss using K2/icefall ( https://github.com/k2-fsa/icefall ; K2 is next generation of Kaldi). As such, it should in principl
210.
by woodson 4y ago
This issue is not restricted to installing packages into your distro-managed Python environment. I had Poetry-managed environments that, after upgrading to poetry>=1.2.0 stopped working (e.g., I couldn't run "poetry update&quot

More ›