LANGUAGE Speech Hearing Auditory Ear Neurosci

<p class="MsoNormal" >CHANG, Edward, Human Speech Cortex, June 1 2015</p>

<p>

This is at 5:15PM in the MBC grad student training program series (Mondays.)

This is at Stanford in Math Corner.

He is a Prof of Neurosurgery at UCSF.&nbsp; EC = Edward Chang

he begins.</p>

<p>

1874&nbsp; Karl Wernicke.&nbsp; studied neuro&nbsp; patients&nbsp; (injuries, stroke)</p>

<p>

The Wernicke region is shown in his own (KWs) diagram .

It&nbsp; was in STS = superior temporal sulcus area

that Wernicke called the"sound images of words"</p>

<p>

and the diagram also shows a region b&nbsp; (Broca's area)

&nbsp; in the frontal lobe for&nbsp;" speech movements."</p>

<p>

but what are the representations in those regions?</p>

<p>

and how is the actual coordination of articulation realized?</p>

<p>

so, how, eg, do related&nbsp; phonemes&nbsp; (buh, fuh, etc) go into

phonological encoding and articulation?</p>

<p>

Patients with lesions in Wernicke's area..&nbsp;

&nbsp;not only cannot understand&nbsp; speech but also produce impaired speech&nbsp; (garbled speech generation)</p>

<p>

this is <b>about 7 synapses up from the cochlea</b>.</p>

<p>

see L. Watts&nbsp; box diagram of nuclei involved.</p>

<p>

speech is preferred over non-speech sounds&nbsp; (in terms of neural responses)&nbsp; by fMRI.</p>

<p>

so, what are the representations involved...</p>

<p>

there are <b>35 to 40&nbsp; phonemes in English</b> that are common&nbsp; (and only 12 in Hawaian)</p>

<p>

so, there is a certain small inventory of phonemes in every language&nbsp;

that produces the whole language... like a a periodic table.</p>

<p>

he shows the International&nbsp; phonetic alphabet</p>

<p>

x axis is the&nbsp; <b>PLACE of&nbsp; generation</b>&nbsp; (bilabial, labiodental, dental, alveolar,

&nbsp;post alveolar, retroflex, palatal, velar, uvular pharyngeal, glottal)</p>

<p>

and y axis&nbsp; is&nbsp; the degree of occlusions (or&nbsp; sound)

plosive,&nbsp; nasal,&nbsp; trill, tap or flap, frictive,&nbsp; lateral fricative,&nbsp; approximant, lateral approximant.</p>

<p>

&nbsp;

so...&nbsp; the fundamental units of phonological reps are NOT segments (phonemes)

but FEATURES,&nbsp; the members of a small category which combine to form the speech sounds of human language.</p>

<p>

FEATURE THEORY:&nbsp; languages do not vary without limit,

&nbsp;but reflect a single general pattern...

(old theory &nbsp;from Chomsky and others)</p>

<p>

EC's&nbsp; group uses ECoG)</p>

<p>

implanted subdural&nbsp; trode arrays for localization of epileptic foci usu for&nbsp; 7 to 10 days.</p>

<p>

usual ECog arrays are spearated 1 cm apart but

he uses&nbsp; 16 by 16 = 256&nbsp; (4 mm trode distance; that is, higher resolution)</p>

<p>

shows it goes v fast... in video&nbsp; he shows real-time activation&nbsp; around STS...

this very fast; so fMRI is way too slow)

paper is Mesgarani, Cheung, Johnson and Chang in Science in 2014...</p>

<p>

STG = sup temp gyrus&nbsp; = Wernicke's.</p>

<p>

he illustrates with the phrase&nbsp; "and what eyes they were"</p>

<p>

he shows location of&nbsp; high gamma response...</p>

<p>

he shows, eg,&nbsp; one trode&nbsp; with&nbsp; all&nbsp; 35 or so phonemes...</p>

<p>

this is LISTENING...</p>

<p>

sibilant fricatives, eg, use as a specialist&nbsp; one trode (and one location)

buh, duh, guh in another&nbsp; trode.</p>

<p>

he shows graph of

trodes&nbsp; (top 100) on x axis

and&nbsp; all 35 phonemes on y axis.</p>

<p>

we did unsupervised clustering on this...</p>

<p>

shows a dendrogram...</p>

<p>

result:&nbsp; consonants were together&nbsp; and&nbsp; sonorants&nbsp; (vowels)&nbsp; were grouped separately.</p>

<p>

consonants grouped as plosives and fricatives&nbsp; and

vowels&nbsp; were&nbsp; low-back,&nbsp; low-front&nbsp; high-front,&nbsp; nasal.</p>

<p>

next graph shows features on the y axis&nbsp; (like dorsal, etc)</p>

<p>

so,&nbsp; we have selectivity to&nbsp; phonetic features...</p>

<p>

STRF&nbsp; spectrotemporal&nbsp; receptor field...&nbsp;

he shows&nbsp; eg plosives versus fricatives for consonants and

vowels in low-back, low front, high front nasal.</p>

<p>

he shows&nbsp; 300 msec record with&nbsp; y axis as&nbsp; freq&nbsp; from 100 hz to&nbsp; 8 khz.</p>

<p>

shows&nbsp; vowel formants&nbsp; (freq cues that help distinguish vowels)</p>

<p>

shows&nbsp; F2&nbsp; (x axis) vs F1&nbsp; on y axis.</p>

<p>

huge overlap</p>

<p>

ascending auditory is tonotopic...&nbsp; each&nbsp; is tuned to partic freq band...&nbsp; (freq decompos in cochlea.</p>

<p>

nex&nbsp; shows&nbsp; group of trodes that have strong response to high gamma

&nbsp;for F1&nbsp; versus&nbsp; another group for F2.</p>

<p>

and shows spectal integration, differential encoding for vowel encoding.</p>

<p>

this is work by Patrick Hullet&nbsp; (in review)</p>

<p>

bSTM = best&nbsp; Spectrotemporal modulation</p>

<p>

that is&nbsp; high order analog of the best freq tuning.</p>

<p>

this is one&nbsp; single&nbsp; parameter that captures the STRF.</p>

<p>

he shows&nbsp; 2d&nbsp; Fourier transforms.</p>

<p>

Q&nbsp; (Tony N)&nbsp; how similar are these tuning curves to&nbsp; single neurons

(is this just spatial averaging)...</p>

<p>

this is "Modulotopic"&nbsp; in organization.</p>

<p>

up to A1&nbsp; is&nbsp; pure tone selectivity&nbsp; but&nbsp; this is not organized by tone.</p>

<p>

primary auditory&nbsp; cortex is embedded inside&nbsp; Sylvian fissue&nbsp; (that's Heschel's gyrus)

he's had two cases where he need to go inside there&nbsp; (via incisions)</p>

<p>

post-doc&nbsp; Liberty Hamilton did this.</p>

<p>

he's been talking abt the features involved.</p>

<p>

experimental: phoneme restoration..</p>

<p>

eg, the words "factor versus faster..."</p>

<p>

eg&nbsp; someone coughs in mid&nbsp; word... but u can fill in the missing phoneme...

(this is called phonetic resoration)</p>

<p>

he plays noise in the middle of the word...</p>

<p>

did pt&nbsp; perceive&nbsp; faster or factor...&nbsp; shows single trode response;

or&nbsp; novel vs&nbsp; nozzle...</p>

<p>

audience comment:&nbsp; top down&nbsp; restoration (by a top down language model)</p>

<p>

we hear words&nbsp; (not phonemes)</p>

<p>

brain activity before the&nbsp; noise predicts what you will hear

&nbsp; (that trajectory is influential in perception)</p>

<p>

conclusion:&nbsp; we've ID'd some important features that we use to perceive words.</p>

<p>

this is all continuous natural speech;</p>

<p>

aud Q...&nbsp; does&nbsp; sequence of presentaiton contribute to this&nbsp; (as opposed to phonmes in isolation)</p>

<p>

For example, consider&nbsp; minority&nbsp; vs&nbsp; tmiroy&nbsp; (some anagram/ permutaiton of minority) processed same or not? ie what is role of context?</p>

<p>

Q&nbsp; how&nbsp; abt people imagining speaking... (is that more like perception or production)

A&nbsp; hard to do that expt...</p>

<p>

(he says... we are&nbsp; far from being&nbsp; able to decode speech)</p>

<p>

Q&nbsp; abt&nbsp; going up or down with pitch...

A&nbsp; pitch IS important&nbsp; (to speaker identity, eg&nbsp; M vs F)&nbsp; also prosidy.</p>

<p>

relative pitch &nbsp;(can be differrent trodes)..</p>

<p>

he was in Mike Merznick's lab and was interested in plasticity and speech&nbsp; (so, did that first)</p>

<p>

in neurosurg: 'we need to know where the eloquent language areas are (so we can avoid them)&nbsp; and where the epileptic foci are.

&nbsp;</p>