Reconstructing intelligible speech from the human auditory cortex | Hacker News Reader

Reconstructing intelligible speech from the human auditory cortex

naplab.ee.columbia.edu

78 pointsJerry215 comments

This is a fantastic bit of work, and I'm excited to see that this technique can be shown to actually work.

From their paper:

We used invasive electrocorticography (ECoG) to measure neural activity from five neurosurgical patients undergoing treatment for epilepsy as they listened to continuous speech sounds. Two of the five subjects had high-density subdural grid electrodes implanted in the left hemisphere with coverage primarily over the superior temporal gyrus (STG), and four of the five subjects had depth electrodes with coverage of Heschl’s gyrus (HG). All subjects had self-reported normal hearing. Subjects were presented with short continuous stories spoken by four speakers (two females, total duration: 30 minutes). To ensure that the subjects were engaged in the task, the stories were randomly paused, and the subjects were asked to repeat the last sentence. The test data consisted of continuous speech sentences and isolated digit sounds. We used eight sentences (40 seconds total) to evaluate the objective quality of the reconstruction models. The sentences were repeated six times in random order, and the neural data was averaged over the six repetitions to reduce the effect of neural noise on comparison of reconstruction models (see Supp. Fig. 1 for the effect of averaging). The digit sounds were used for subjective intelligibility and quality assessment of reconstruction methods and were taken from a publicly available corpus, TI-46. We chose 40 digit sounds (zero to nine), spoken by four speakers (two females) that were not included in the training of the models. Reconstructed digits were used as the test set to evaluate subjective intelligibility and quality of the models.

Do not expect this to generalize easily to a noninvasive hairnet: ECoG is so, so much nicer than EEG from a signal to noise perspective, not to mention latency and localization. The only drawback is you have to crack someone's skull to do it...

> Do not expect this to generalize easily to a noninvasive hairnet: ECoG is so, so much nicer than EEG from a signal to noise perspective, not to mention latency and localization. The only drawback is you have to crack someone's skull to do it...

Do you know if these implanted electrods interfere with transcranial EEG measurement? If not, is anyone gathering data from these invasively-instrumented humans to see if it's possible to learn a mapping between the two? It'd be a long shot, but still...

it's easier to just replicate the work with EEG and see how well you do. Having worked with these signals, my guess is extremely poorly.

> The only drawback is you have to crack someone's skull to do it...

Couldn't the recent work in implanting mesh electronics via syringe "solve" this problem?

well, you’ll have a smaller hole then. Epidural ECoG would be marginally safer as well; not sure which they used here, but probably subdural.

@ML experts: could a model be trained from ECoG data so that we can better isolate signals from EEG?

Can this method be used to extract speech by people who subvocalize, that is, where auditory processing is going on in the absense of speech by another? Also, this confirms to me that we are merely, or at least mostly, made of electrified meat.

Can imagine this having really profound implications for folks with ALS, etc down the line. Really impressive.

Interesting work.

This proves that the Borg's collective voice uses " DNN + Spectrogram" to synthesize their voice from the drones auditory cortex, compare: https://www.youtube.com/watch?v=ql83z8yBx2M&feature=youtu.be...

Here's a link to the research paper itself: https://www.nature.com/articles/s41598-018-37359-z.pdf

I wonder why they used the WORLD vocoder and whether quality can be improved by using other vocoders like wavenet or lpcnet.

That would be a nice project idea, exchanging the vocoder with something like wavenet and making it end-to-end.

The cool thing is that they published here (http://naplab.ee.columbia.edu/naplib.html) all their training data and code, so that should not be actually difficult.

Edit: My mistake, the data is not available, but they say in the paper that it is available under request..

Can someone summarize what mathematical operations are used in the reconstruction? E.g. is there an (inverse) Fourier transform? (Asking since the first thing the ear does is basically to perform a Fourier transform, separating the auditory signal in different frequencies).

Or do they use a more opaque ML approach?

I just skimmed over the paper (don't really have time to sit and read it down right now), but it seems it's about like this:

1. They captured the ECoG of people hearing stories (for training). They also captured ECoG of people saying numbers 0 to 9.

2. They trained the vocoder with those parameters as an autoencoder that goes from 516 parameters down to 256 and back again to 516. The 516 parameters are spectral envelope, pitch, voicing and aperiodicity.

3. They trained the models to receive as input the HG and LF signals (from the ECoG I think -- don't really know wht these names mean, just what the figures on the paper show). The (expected) output for the models was either the audio spectrogram or the vocoders 256 encoding.

So I'd go with "opaque ML approach" to answer your last question. As for the mathematical operations involved (in general), I dind't read the detailed explanation (if there is one) of how the networks are working, but probably ReLU activations and summation over parameters, like most DNNs.

Now can it go the other way?

Reply on news.ycombinator.com