Fermion Research

Reading the neurons of a speech model

We opened up Phonon-2.5, the speech recognition model inside Detta, and found it keeping careful track of things nobody ever asked it about.

If you play Phonon-2.5 a few seconds of a dog barking, it usually writes nothing, which is correct, because dogs rarely say anything worth transcribing. If you play it hundreds of clips with nobody in them, such as coughs, rain, a creaking door or a passing helicopter, then about one time in fifteen it writes something anyway. Sometimes it is "Hey", sometimes "Bye." and sometimes a hopeful "Ja?". Once, from a clip in which nobody was speaking, it wrote "Mercedes West Mercedes West" with the calm confidence of a court stenographer.

Our model is one of the well-behaved ones. Whisper-small, a widely used open model, writes something on 98.6% of these clips, from a lone "you" to "2.5 km" to a long loop of the Georgian letter ლ. People do a version of this too, seeing faces in power sockets and hearing their name in the hiss of a shower, and psychologists call it pareidolia. What surprised us is how literal the machine version is. Inside Phonon-2.5 there is a direction that roughly means "somebody is talking", and on these clips it lights up, so inside, the model responds as if a person were there. We found that spot, and we found a way to turn it down at a cost to real speech that mostly rounds to zero.

On the way we kept tripping over other surprises, because a model trained to do one thing, turning sound into words, turns out to keep notes on almost everything else. It keeps one neuron for each of its 25 languages, and it can tell a kitchen from the inside of a car. This post is a tour of those notes, and it ends with the ghost.

How you open up a speech model

Phonon-2.5 has an encoder, which turns audio into a stream of internal numbers, and a decoder, which turns that stream into text. The encoder is a stack of 24 layers, and it helps to picture them as the floors of a building that the sound climbs, becoming a little more abstract on each floor. Every floor holds thousands of small dials, usually called neurons, that rise and fall as the audio plays. Nobody tells them what to mean, because the model is only ever scored on getting the words right, so we went floor by floor asking a simple question. Is there a single dial that goes up for Greek and stays down for every other language, or one for whispering, or one for being in a car?

Searching thousands of dials for one that happens to match a label is a great way to fool yourself, so every dial in this post was chosen on one set of recordings, tested on completely different speakers, rooms and microphones, and had to clearly beat the luckiest dial we could find with the labels scrambled.1 We ran the same search on two public models, parakeet-tdt-0.6b-v3 and Whisper-small, to see what was special to ours and what seems true of speech models in general.

Figure 1. The room is noticed near the input and fades by the top. The language is known by block 6 but only written into the transcript at blocks 18 to 22, and the direction behind phantom voices lives upstairs.

One neuron for every language

Phonon-2.5 transcribes 25 languages and is never told which one it is hearing, yet for every one of them there is a single neuron that lights up for that language and stays quiet for the other 24. Models like Whisper are trained to name the language outright before they transcribe, but Phonon-2.5, like Parakeet, is never asked to, and the only explanation on offer has been that it picks the language up implicitly, somewhere in the network. In ours, that somewhere is one neuron per language. Most of these neurons sit in the first third of the encoder. We then played the same neurons European Parliament speeches they had never heard, and they named the language correctly 99 times in 100, better than a model built to do nothing but identify languages.2

Figure 2. A single neuron per language beats a dedicated language-identification model. The much bigger reader collapses because it memorised the benchmark's recordings instead of the languages.

This figure is interactive and needs JavaScript.

Figure data for agents: One neuron per language (generated from neurons.json)

Each row is the neuron that goes with one of the 25 languages. Pick a recording and see which row lights up as it plays.

The neurons (one row each in the figure; B = encoder block, ff = feed-forward layer, N = unit):

LanguageNeuronEncoder block
EnglishB16.ff2.N210016
GermanB10.ff2.N391610
SpanishB6.ff2.N14916
FrenchB4.ff2.N20884
ItalianB6.ff2.N26036
PortugueseB6.ff2.N9486
DutchB6.ff2.N4046
PolishB7.ff1.N40417
CzechB18.ff2.N43918
SlovakB18.ff2.N119218
RussianB17.ff2.N27617
UkrainianB18.ff2.N21318
BulgarianB7.ff2.N12257
CroatianB17.ff2.N267417
SlovenianB18.ff2.N39018
LatvianB7.ff2.N4607
LithuanianB7.ff2.N347
EstonianB6.ff2.N15656
FinnishB7.ff2.N557
SwedishB7.ff2.N11317
DanishB7.ff2.N3347
HungarianB7.ff2.N27607
RomanianB7.ff2.N11667
GreekB7.ff1.N13937
MalteseB10.ff2.N48110

For each clip, the neuron that fires hardest on average over the clip: right on 25 of 25 clips.

ClipStrongest neuron over the clipVerdictClip source
EnglishEnglishrightFLEURS test clip 1904 (English), CC BY 4.0: 3.8 s of continuous speech from 2.0 s in, level-matched.
GermanGermanrightFLEURS test clip 2003 (German), CC BY 4.0: 3.9 s of continuous speech from 6.5 s in, level-matched.
SpanishSpanishrightFLEURS test clip 2001 (Spanish), CC BY 4.0: 3.1 s of continuous speech from 1.9 s in, level-matched.
FrenchFrenchrightFLEURS test clip 1857 (French), CC BY 4.0: 3.6 s of continuous speech from 0.9 s in, level-matched.
ItalianItalianrightFLEURS test clip 1787 (Italian), CC BY 4.0: 3.9 s of continuous speech from 3.6 s in, level-matched.
PortuguesePortugueserightFLEURS test clip 1960 (Portuguese), CC BY 4.0: 3.2 s of continuous speech from 13.8 s in, level-matched.
DutchDutchrightFLEURS test clip 1927 (Dutch), CC BY 4.0: 3.1 s of continuous speech from 0.4 s in, level-matched.
PolishPolishrightFLEURS test clip 1829 (Polish), CC BY 4.0: 3.5 s of continuous speech from 2.8 s in, level-matched.
CzechCzechrightFLEURS test clip 1970 (Czech), CC BY 4.0: 3.2 s of continuous speech from 6.1 s in, level-matched.
SlovakSlovakrightFLEURS test clip 1781 (Slovak), CC BY 4.0: 3.2 s of continuous speech from 3.1 s in, level-matched.
RussianRussianrightFLEURS test clip 1917 (Russian), CC BY 4.0: 3.2 s of continuous speech from 2.0 s in, level-matched.
UkrainianUkrainianrightFLEURS test clip 1827 (Ukrainian), CC BY 4.0: 3.7 s of continuous speech from 0.6 s in, level-matched.
BulgarianBulgarianrightFLEURS test clip 1960 (Bulgarian), CC BY 4.0: 3.8 s of continuous speech from 1.3 s in, level-matched.
CroatianCroatianrightFLEURS test clip 1776 (Croatian), CC BY 4.0: 3.2 s of continuous speech from 0.8 s in, level-matched.
SlovenianSlovenianrightFLEURS test clip 1738 (Slovenian), CC BY 4.0: 3.1 s of continuous speech from 3.2 s in, level-matched.
LatvianLatvianrightFLEURS test clip 1951 (Latvian), CC BY 4.0: 3.0 s of continuous speech from 3.3 s in, level-matched.
LithuanianLithuanianrightFLEURS test clip 1822 (Lithuanian), CC BY 4.0: 3.1 s of continuous speech from 0.2 s in, level-matched.
EstonianEstonianrightFLEURS test clip 1831 (Estonian), CC BY 4.0: 3.2 s of continuous speech from 9.5 s in, level-matched.
FinnishFinnishrightFLEURS test clip 1862 (Finnish), CC BY 4.0: 3.1 s of continuous speech from 7.0 s in, level-matched.
SwedishSwedishrightFLEURS test clip 1844 (Swedish), CC BY 4.0: 3.6 s of continuous speech from 0.7 s in, level-matched.
DanishDanishrightFLEURS test clip 1842 (Danish), CC BY 4.0: 3.2 s of continuous speech from 0.4 s in, level-matched.
HungarianHungarianrightFLEURS test clip 1874 (Hungarian), CC BY 4.0: 3.1 s of continuous speech from 4.0 s in, level-matched.
RomanianRomanianrightFLEURS test clip 1999 (Romanian), CC BY 4.0: 3.0 s of continuous speech from 9.9 s in, level-matched.
GreekGreekrightFLEURS test clip 1993 (Greek), CC BY 4.0: 3.0 s of continuous speech from 0.9 s in, level-matched.
MalteseMalteserightFLEURS test clip 1920 (Maltese), CC BY 4.0: 3.0 s of continuous speech from 7.2 s in, level-matched.

On these 25 clips, a few seconds of continuous speech each, the strongest of the 25 neurons names the right language 25 times; on all 240 such windows from the same FLEURS test sets it is right 94% of the time. Each row is scaled to its own spread over all the windows.

Interactive. Each column is 80 milliseconds of a public FLEURS test clip.

It carries a speedometer that nobody installed

Speed a recording up to two and a half times normal speed and Phonon-2.5 struggles badly, yet inside it tracks how fast you are talking. Partway up the model there is one direction whose strength rises the faster the audio goes, and across our test clips its strength and the real speed move together almost perfectly. We turned it into a speedometer the way you would calibrate a real one: we sped practice clips up by known amounts, 1×, 1.5×, 2× and 2.5×, measured that direction on each, and drew a straight line through the points, so that every strength corresponds to a speed. For a new recording we measure the direction over the clip and read the speed off that line; the clip in the figure below reads 1.25× at normal speed, 1.71× at 1.5×, 2.10× at 2× and 2.55× at 2.5×.

Then we act on it. When the speedometer says a clip is clearly fast, we slow the audio down by exactly that much before transcribing, so a reading of 2.55× means playing it at 1/2.55 of the speed, and normal clips stay below the cut-off and are left alone; only 1 of 60 normal-speed clips was flagged. On test clips the speedometer had never seen, that takes errors on 2.5× speech from 23.6 per hundred words to 3.2.

Figure 3. One neuron at block 12 reads how fast someone is talking. It flags almost no normal-speed clips and every clip played at double speed or faster.

This figure is interactive and needs JavaScript.

Figure data for agents: Read the speedometer, then slow down (generated from tempo.json)

What was said: “When did you come, Bartley, and how did it happen? You haven't spoken a word.” (LibriSpeech test-clean 4446-2275-0008 (CC BY 4.0), sped up with sox, pitch kept). A clip is slowed down when the speedometer reads above 1.3×.

Playback speedSpeedometer readsActionPhonon-2.5 writes as isPhonon-2.5 writes after slowing down
1×1.25×normal speed, left alone“When did you come, Bartley, and how did it happen? You haven't spoken a word.” (every word right)not slowed; transcript stays the same
1.5×1.71×too fast, slowed down 1.71דWhen did you come, Bartley, and how did it happen? You haven't spoken a word.” (every word right)“When did you come, Bartley, and how did it happen? You haven't spoken a word.” (every word right)
2×2.10×too fast, slowed down 2.10דWhy did you come, Bartley, and how did it happen? You haven't spoken a word.” (wrong: Why)“When did you come, Bartley, and how did it happen? You haven't spoken a word.” (fixed by slowing down: When)
2.5×2.55×too fast, slowed down 2.55דYou haven't spoken a word.” (missing: When did you come Bartley and how did it happen)“When did you come, Bartley, and how did it happen? You haven't spoken a word.” (fixed by slowing down: When did you come, Bartley, and how did it happen?)
3×2.47×too fast, slowed down 2.47דWhat do you come or do happen?” (wrong: What do or do; missing: Bartley and how You haven't spoken a word)“When did you come, Bartley, and how did it happen? You haven't spoken a word.” (fixed by slowing down: When did Bartley, and how did it You haven't spoken a word.)

Across 60 test clips per speed (words wrong per 100):

Playback speedErrors per 100 words as isAfter reading the speedometer and slowing downClips flagged as fast
1×2.62.61 of 60
1.5×3.42.755 of 60
2×5.83.160 of 60
2.5×23.63.260 of 60
3×66.14.960 of 60
Interactive. Reading the speedometer and slowing down repairs time-stretched speech almost completely.

It knows what room you are in

A speech model hears more than the words. We played the same spoken sentences over background recordings from 17 different places, from a kitchen to the inside of a moving car, and looked for a neuron that fires when one particular place is in the background and stays quiet for all the others. We picked each neuron on one batch of recordings, then tested it on new speakers and new recordings it had never heard. For 15 of the 17 places a single neuron does this reliably; the kitchen neuron, for example, tells kitchen recordings apart from the rest almost perfectly. So when we say the model knows which room you are in, we mean that the neuron that fires hardest names the place, and nobody built it to do that. The room is loudest in the first few layers and fades on the way up: a simple reader can name the place 86% of the time from the first layer but only 16% of the time from the last, so the model notices where you are and then sets it aside to concentrate on the words. Two public speech models, Parakeet and Whisper-small, keep track of most of the same things, so this looks like something speech models learn in general.

This figure is interactive and needs JavaScript.

Figure data for agents: It knows what room you are in (generated from neurons.json)

The same sentence, read once, mixed over recordings of six places. Each row is the neuron that picked out one of them, all in the encoder’s first three blocks.

The neurons (one row each in the figure; B = encoder block, ff = feed-forward layer, N = unit):

PlaceNeuronEncoder block
KitchenB2.ff1.N36962
Inside a carB0.ff1.N2300
Train stationB1.ff1.N40601
ParkB2.ff1.N25732
RestaurantB2.ff1.N10252
CafeteriaB1.ff1.N23931

For each clip, the neuron that fires hardest on average over the clip: right on 6 of 6 clips.

ClipStrongest neuron over the clipVerdictClip source
No room noiseKitchenno place to find (clean speech)LibriSpeech test-clean 1284-1180-0009, CC BY 4.0.
KitchenKitchenrightLibriSpeech test-clean 1284-1180-0009, CC BY 4.0, mixed with DEMAND DKITCHEN (CC BY-SA 3.0) at 5 dB; this mix is shared under CC BY-SA 3.0.
Inside a carInside a carrightLibriSpeech test-clean 1284-1180-0009, CC BY 4.0, mixed with DEMAND TCAR (CC BY-SA 3.0) at 5 dB; this mix is shared under CC BY-SA 3.0.
Train stationTrain stationrightLibriSpeech test-clean 1284-1180-0009, CC BY 4.0, mixed with DEMAND PSTATION (CC BY-SA 3.0) at 5 dB; this mix is shared under CC BY-SA 3.0.
ParkParkrightLibriSpeech test-clean 1284-1180-0009, CC BY 4.0, mixed with DEMAND NPARK (CC BY-SA 3.0) at 5 dB; this mix is shared under CC BY-SA 3.0.
RestaurantRestaurantrightLibriSpeech test-clean 1284-1180-0009, CC BY 4.0, mixed with DEMAND PRESTO (CC BY-SA 3.0) at 5 dB; this mix is shared under CC BY-SA 3.0.
CafeteriaCafeteriarightLibriSpeech test-clean 1284-1180-0009, CC BY 4.0, mixed with DEMAND PCAFETER (CC BY-SA 3.0) at 5 dB; this mix is shared under CC BY-SA 3.0.

Compared among these six places, the right neuron is the strongest for all six. The noise is from the public DEMAND recordings, mixed so that the speech carries about three times the power of the noise.

Interactive. The sentence is a public LibriSpeech recording and the places come from the public DEMAND recordings. Pick a place and the row for it should light up.

It can hear a stutter

The same method can be pointed at what a voice reveals about the person speaking. A neuron chosen on one podcast still finds stuttered speech on a completely different one, because a stutter is something you can actually hear. These are research findings from public collections, not medical tests.

Sometimes it hears a voice that is not there

This brings us back to Mercedes West. Inside the model there is something that works like a dial labelled “someone is talking right now”. When it is turned up the model writes words, and when it is down the model writes nothing. Normally only a real voice turns it up, but some sounds, like a cough, rain or a slammed door, nudge it up by mistake, and the model writes words nobody said. We found the dial by taking clips with nobody talking and sorting them into two piles, the ones where the model wrote something anyway and the ones where it rightly stayed silent. Looking inside the model, the difference between the two piles was that dial: it was turned up on the phantom pile. When the model writes “Bye.” over an empty clip, its insides behave as if it heard a person, the way you can see a face in a cloud.

To prove it, we did the reverse on purpose. We took what the inside of the model looks like when it hears a real word, say “mother”, and planted that pattern into four seconds of pure silence. For each of the eight words we tried, the model wrote the word 40 times out of 40, while random tweaks of the same size never did. “Without” came out as “without without”, and “mother” came out as “Mother”, with a capital, as if someone had turned to speak to her. The model writes whatever its insides say it heard, even over silence, and that is all a phantom is. We had made a ghost on purpose.

Figure 4. On 2,000 ESC-50 recordings of everyday sounds with nobody talking, Whisper-small writes something on 1,971 and Phonon-2.5 on 129. On the first five clips Whisper-small writes words and Phonon-2.5 writes nothing, which is the right answer; on the last two Phonon-2.5 writes a phantom and the bouncer, introduced later in this post, removes it.

Whole sentences can be planted in silence

One planted word made us wonder how far the trick goes, so we wrote down twenty sentences before running anything, such as “There is nobody here” and “Is anybody there”, and built each one from the model’s own memory of speech. For every word piece we took Phonon-2.5’s average state at layer 18 in the moments it writes that piece in ordinary recordings, and added those states one after another into a second or two of audio in which every sample is zero. Eighteen of the twenty came out word for word. We never planted punctuation, so the question marks in “Is anybody there?” and “Who is speaking?” are the model’s own. Random vectors of the same sizes, thirty for each sentence, never wrote a word of it in 3,990 tries.

All of this runs on a research build of the model; Detta itself skips digital silence before the model ever runs. Noise, we think, does a small and clumsy version of the same push by accident.

This figure is interactive and needs JavaScript.

Figure data for agents: Sentences planted in silence (generated from planted.json)

Each input file is digital silence (every sample zero); inside the model one vector per word piece is added. 18 of the 20 sentences, written down before anything was run, came out word for word; random vectors of the same sizes wrote no word of any of them in 3,990 clips.

Planted sentencePhonon-2.5 writesWord for word?Vectors addedSeconds of silenceRandom-vector controls
is anybody thereIs anybody there?yes51.4530 tried, 0 wrote it
there is nobody hereThere is nobody here.yes61.6130 tried, 0 wrote it
i can hear youI can hear you.yes51.4530 tried, 0 wrote it
can you hear meCan you hear me?yes51.4530 tried, 0 wrote it
nobody said anythingNobody said anything.yes71.7730 tried, 0 wrote it
this file is silentcould not be built---the word “file” appears too rarely in our recordings to build
hello worldHellono41.2930 tried, 0 wrote it
good morningGood morning.yes31.1330 tried, 0 wrote it
thank you very muchThank you very much.yes51.4530 tried, 0 wrote it
i am not hereI am not here.yes41.2930 tried, 0 wrote it
please do not wake mePlease do not wake me.yes71.7730 tried, 0 wrote it
the house was emptyThe house was empty.yes71.7730 tried, 0 wrote it
listen to the silenceListen to the silence.yes61.6130 tried, 0 wrote it
who is speakingWho is speaking?yes51.4530 tried, 0 wrote it
there was no one in the roomThere was no one in the room.yes81.9330 tried, 0 wrote it
something is in hereSomething is in here.yes51.4530 tried, 0 wrote it
say somethingsay something.yes31.1330 tried, 0 wrote it
i hear voicesI hear voices.yes61.6130 tried, 0 wrote it
all is quietAll is quiet.yes41.2930 tried, 0 wrote it
nothing to see hereNothing to see here.yes51.4530 tried, 0 wrote it
Interactive. Every sample of the file is zero.

The ghost needed a bouncer

The obvious fix was to hold that “someone is talking” dial down, which we call the clamp: deep in the model, whenever the dial rose above its usual level, we trimmed it back. Two thirds of the clips with phantom words came out clean, but real speech paid for it, and Greek transcripts got 3.5 more words wrong in every hundred. The clamp could not tell a cough from a quiet person in a noisy room, because with the clamp on all the time, it also turned the dial down while real people were talking, and their words got garbled or dropped.

So we hired a bouncer: a tiny checker, about 10 KB, lower in the model, that decides moment by moment whether anyone is really speaking. The clamp is only allowed to act when the bouncer says the room is empty, so the phantoms go and real speech is left alone. On the recordings we built it with, it worked beautifully. On fresh recordings of real people talking in real noise, it kept turning them away at the door, because everything it knew about speech had come from clean recordings, and it let the clamp loose on almost 9% of the moments in which those people were actually talking.

Figure 5. Each mistake here lets the clamp loose on real words. The original bouncer had only seen clean recordings and made this mistake most often on home microphones and in real noise; after retraining it almost never does.

The fix was to retrain the bouncer on hard, messy, real speech, including air traffic radio and distant microphones in meetings.3 The retrained gate removed 87% of the phantom words on clips it had never seen, left real speech essentially untouched across ten sets of recordings, and costs less than 1% of the encoder's running time. We also tried the obvious shortcut of letting Silero, the standard open voice detector, decide when the clamp may act. It removed about as many phantoms, but it let the clamp loose in the middle of real speech and changed 42 of 100 clips of people talking, where our bouncer changed none. The best speech detector for this job turned out to be the one already inside the model.

Figure 6. Each row is a fresh set of real speech. The original gate cost noisy speech the most, while the retrained gate sits on zero almost everywhere.

This figure is interactive and needs JavaScript.

Figure data for agents: What the model writes when nobody is talking (bouncer off and on) (generated from bouncer.json)

Research copy of Phonon-2.5. Of 150 public clips with nobody speaking, 150 get phantom words with the bouncer off and 5 with it on; of 100 clips of real speech, 100 come out word for word the same with the bouncer on (0 changed).

The clips you can play in the figure:

Sound (nobody speaking)LengthBouncer off: the model writesBouncer on: the model writesNoteSource
A big dog barking0.9 s“What?”nothingFreesound 182395 by pgonsilva, CC BY 3.0, via FSD50K
A rooster crowing4.8 s“Ha ha”nothingFreesound 200339 by jsbarrett, CC0, via FSD50K
One short note on a French horn1.3 s“Bye.”nothingFreesound 373240 by Samulis, CC0, via FSD50K
A wompoo fruit dove1.4 s“Boa, pois.” (Portuguese, roughly “good, then”)nothingFreesound 114449 by soundbytez, CC BY 3.0, via FSD50K
A clarinet note with tremolo4.5 s“Uh”nothingFreesound 356501 by MTG, CC BY 3.0, via FSD50K
A glass bottle being opened4.0 s“Давай.” (Russian, “come on”)nothingFreesound 188903 by plingativator, CC BY 3.0, via FSD50K
A zipper0.6 s“What?”nothingFreesound 195070 by punpcklbw, CC BY 3.0, via FSD50K
A single cat meow0.8 s“Bill”nothingFreesound 334433 by ebcrosby, CC BY 3.0, via FSD50K
A double bass1.8 s“I'm telling you, I'm telling you”nothingFreesound 372332 by SSS_Samples, CC BY 3.0, via FSD50K
A bunch of keys5.2 s“Cool.”nothingFreesound 127558 by romulofs, CC0, via FSD50K
A knock on a door4.2 s“Hmm”nothingFreesound 193829 by Ligidium, CC0, via FSD50K
A clock ticking on a table4.0 s“Sono davvero molto presenta.” (Italian, roughly “they really are very present”)“Sono davvero”only partly cleanedFreesound 388903 by ycbcr, CC BY 3.0, via FSD50K
A green-winged macaw1.3 s“когда” (Russian, “when”)“когда”one of the clips the bouncer lets throughFreesound 111075 by soundbytez, CC BY 3.0, via FSD50K

Every clip in the check (one dot each in the figure):

#KindBouncer offBouncer on
1Everyday sound“Uh”nothing
2Everyday sound“когда”“когда”
3Everyday sound“Sorry to say.”nothing
4Everyday sound“Boa, pois.”nothing
5Everyday sound“Не прошло.”nothing
6Everyday sound“Если усе, десу снять.”nothing
7Everyday sound“Ему..”nothing
8Everyday sound“Cool.”nothing
9Everyday sound“Well, yeah.”nothing
10Everyday sound“What?”nothing
11Music“Ah.”nothing
12Music“Да.”nothing
13Music“Tā.”nothing
14Music“Да.”nothing
15Music“Меру”nothing
16Music“Nem”nothing
17Everyday sound“Go”nothing
18Everyday sound“What?”nothing
19Music“Uh”nothing
20Everyday sound“Давай.”nothing
21Everyday sound“Hmm”nothing
22Everyday sound“What?”nothing
23Everyday sound“Ha ha”nothing
24Everyday sound“Wow”nothing
25Everyday sound“What?”nothing
26Everyday sound“Que?”nothing
27Everyday sound“What?”nothing
28Everyday sound“What?”nothing
29Everyday sound“Пошли.”nothing
30Music“Mm-hmm.”nothing
31Music“Fras”nothing
32Music“Хм”nothing
33Music“А ну”nothing
34Music“I would say”nothing
35Everyday sound“Fight”nothing
36Music“Qua igual igual vo.”nothing
37Everyday sound“Bill”nothing
38Everyday sound“Ele vai.”nothing
39Everyday sound“Слышу, да?”nothing
40Music“То”nothing
41Music“Yeah.”nothing
42Music“Uh”nothing
43Music“Yeah.”nothing
44Music“Yeah, yeah, yeah, yeah.”nothing
45Music“Uh”nothing
46Music“Okay”nothing
47Music“I'm telling you, I'm telling you”nothing
48Music“Bom”nothing
49Music“Yeah.”nothing
50Music“Yeah.”“Okay.”
51Music“No.”nothing
52Music“Bye.”nothing
53Music“Qua?”nothing
54Music“Um”nothing
55Music“Um”nothing
56Music“Yep.”nothing
57Music“No.”nothing
58Everyday sound“No?”nothing
59Everyday sound“Okay.”nothing
60Everyday sound“Sono davvero molto presenta.”“Sono davvero”
61Everyday sound“Oh”nothing
62Everyday sound“Fair”nothing
63Everyday sound“Зло”nothing
64Everyday sound“No”nothing
65Everyday sound“Tu”nothing
66Music“Так”nothing
67Everyday sound“Да”nothing
68Everyday sound“Ха”nothing
69Everyday sound“Od”nothing
70Everyday sound“Кры”nothing
71Everyday sound“Как ты?”nothing
72Human sound, not speech“No”nothing
73Human sound, not speech“They don't do this to me.”nothing
74Human sound, not speech“Ух.”nothing
75Human sound, not speech“No, no, no, no, no.”nothing
76Human sound, not speech“Wake up wake up.”nothing
77Human sound, not speech“Bless you!”nothing
78Human sound, not speech“Oh”nothing
79Human sound, not speech“So”nothing
80Human sound, not speech“Uh”nothing
81Human sound, not speech“Yeah”nothing
82Human sound, not speech“A fin”nothing
83Human sound, not speech“And I was so scared.”nothing
84Human sound, not speech“How could you?”“How could you?”
85Human sound, not speech“Ha it's got me! Hell!”nothing
86Human sound, not speech“No, please no”nothing
87Human sound, not speech“Mm-hmm.”nothing
88Human sound, not speech“Oh no, no no”nothing
89Human sound, not speech“Ha”nothing
90Human sound, not speech“No”nothing
91Human sound, not speech“Co”nothing
92Human sound, not speech“No hell”nothing
93Human sound, not speech“А”nothing
94Human sound, not speech“Хм”nothing
95Human sound, not speech“Uh”nothing
96Human sound, not speech“Ooh”nothing
97Human sound, not speech“Uh-huh.”nothing
98Human sound, not speech“Wait no”nothing
99Human sound, not speech“Oh”nothing
100Human sound, not speech“Uh yeah”nothing
101Human sound, not speech“Uh ge, oh geah”nothing
102Human sound, not speech“I hear you.”nothing
103Human sound, not speech“Mmm”nothing
104Human sound, not speech“Oh no”nothing
105Human sound, not speech“Uh oh”nothing
106Human sound, not speech“Uh”nothing
107Human sound, not speech“Uh-huh.”nothing
108Human sound, not speech“Uh”nothing
109Human sound, not speech“No”nothing
110Human sound, not speech“Who dangerous?”nothing
111Human sound, not speech“So”nothing
112Human sound, not speech“Jā.”“Yeah.”
113Human sound, not speech“Ah!”nothing
114Human sound, not speech“Com”nothing
115Human sound, not speech“Why are you owning?”nothing
116Human sound, not speech“Oh”nothing
117Human sound, not speech“Hak tu”nothing
118Human sound, not speech“Uuuuh”nothing
119Human sound, not speech“No They don't love me anymore.”nothing
120Human sound, not speech“Doctor”nothing
121Human sound, not speech“Я все”nothing
122Human sound, not speech“Oh no.”nothing
123Human sound, not speech“Oh it isn't they don't love me anymore.”nothing
124Human sound, not speech“Come.”nothing
125Human sound, not speech“Please stop”nothing
126Human sound, not speech“Хрину”nothing
127Human sound, not speech“Send it back.”nothing
128Human sound, not speech“Here”nothing
129Human sound, not speech“Où les couilles nous là”nothing
130Human sound, not speech“Ha ha”nothing
131Human sound, not speech“Don't have”nothing
132Human sound, not speech“What?”nothing
133Human sound, not speech“Oh who are you? Get away from me!”nothing
134Human sound, not speech“Hello”nothing
135Human sound, not speech“All right,”nothing
136Human sound, not speech“I don't know, I don't know.”nothing
137Human sound, not speech“Ah”nothing
138Human sound, not speech“When it's around”nothing
139Human sound, not speech“Nö”nothing
140Human sound, not speech“Anymore. They don't love me.”nothing
141Human sound, not speech“Oh”nothing
142Human sound, not speech“Oh”nothing
143Human sound, not speech“Mom”nothing
144Human sound, not speech“No who are you?”nothing
145Human sound, not speech“Whoa!”nothing
146Human sound, not speech“Ah”nothing
147Human sound, not speech“Mm-mm.”nothing
148Human sound, not speech“No please win go please”nothing
149Human sound, not speech“Um”nothing
150Human sound, not speech“Ha ha”nothing
151Noisy real speech“Punk and crush. And you still don't know why. I my understanding is uh there's only one”same
152Noisy real speech“The result is that, in order to keep up a decent appearance, the former habitually live hand to mouth to a greater extent than the latter,”same
153Noisy real speech“I tried uh I tried to access my personal account in Google in uh Facebook ads and”same
154Noisy real speech“As one review put it, the janitor is the undercover columns.”same
155Noisy real speech“When a draft spoke with that silence, the nephano showed no emotion.”same
156Noisy real speech“Searchers are a lot of people. Eat them that actually would like to have a little more freedom and space to be able”same
157Noisy real speech“De Guuiche, observing that she was alone, near a thicket constructed of painted cloth, approached her,”same
158Noisy real speech“The machine level for for the clusters, they might have been a disadvantage, but that becomes a programming tailwind opportunity.”same
159Noisy real speech“It was no more exciting than an overturned carriage or a tree ranch off the road.”same
160Noisy real speech“Fun. Well Dave, I know it's a short Friday because uh you took a red eye and uh I gotta get back on the AI leaders.”same
161Noisy real speech“I took it up, and it had solved the engine oiling problem, and my walk to Lawrence, like a tramp actor's, was off.”same
162Noisy real speech“I could just see my uncle at cool link from the raft, and Hans still in a stone sitting fire under the action of the U.S. whistle, which I sat”same
163Noisy real speech“She, meanwhile, passed her life with her parents in the strictest retirement, never letting herself be seen, but shunning every eye lest it should weave her misfortune in her face.”same
164Noisy real speech“cars. But the challenge is really, um, the bending over, obviously, but when I got the new studio to put a lift in there”same
165Noisy real speech“Great Mother's Day response, which is in the course of a year. Long journey.”same
166Noisy real speech“centabolic in this phase, like we're gonna do something a little bit different. Like instead of just doing these kind of five by five or a three by five routine, I'm gonna have”same
167Noisy real speech“His pulse beat fast as he recalled the friends and comrades with whom he had spent the last few years in that vanished city.”same
168Noisy real speech“The National Hockey Center features three international-level synthetic pitches.”same
169Noisy real speech“Her condition was yesterday this very was critical but stable.”same
170Noisy real speech“I can see Douglas there before the fire, to which he had got up to present his back, looking down at his interlocutor with his hands in his pockets.”same
171Noisy real speech“He took no notice of her. He looked at me, but as if instead of me he saw what he spoke of.”same
172Noisy real speech“That's a courtesy, that's about friendship, and that's not about letting people know we kept running 150. We lived a life of discipline. My guest, guys, is saying”same
173Noisy real speech“It was fast, it was comfortable. I think if you were driving around town, totally cool. But I have to go, you know.”same
174Noisy real speech“Be careful with your prognostication, said the stranger.”same
175Noisy real speech“Heraldic matters in the Republic of Ireland are handled by the Genealogical Office.”same
176Speech in loud noise“He came down to earth, lived among men, suffered, was crucified, and then he died, standing clearly before us, so that our hearts and eyes may fasten upon him.”same
177Speech in loud noise“Mainhall vouched for her constancy with a loftiness that made Alexander smile, even while a kind of rapid excitement was tingling freedom.”same
178Speech in loud noise“Do we really know the mounty well when we are not acquainted with the cabin?”same
179Speech in loud noise“No harm would have been done had it not been that as he passed your door he perceived the key which had been left by the carelessness of your servant.”same
180Speech in loud noise“The electric light can scarcely penetrate through the dense curtain which is dropped over the theatre on which the Battle of the Elements is about to be waged.”same
181Speech in loud noise“Then of a sudden, at midday, the sun shot out, hot and still, no breath of air stirred, the sky was like blue steel, the earth seemed,”same
182Speech in loud noise“Did you three days? She light the fire, every silver spread and circular.”same
183Speech in loud noise“What is your name, Lordy? asked the little scholar, presently.”same
184Speech in loud noise“Indeed, there were only one or two strangers who could be admitted among the sisters without producing the same result.”same
185Speech in loud noise“There was something individual about the great farm, a most unusual gymnast and careful details.”same
186Speech in loud noise“He might have had that forch check for the face of it, if he'd been sharp,”same
187Speech in loud noise“A narrow channel formed a passage through the ridge of rocks that protected it from the open sea, and which, even in the roughest weather, would ensure the calmness of its waters.”same
188Speech in loud noise“Then they started on again, and two hours later came in sight of the house of doctor Pitt.”same
189Speech in loud noise“You know, I thought began to fix conscious to night. For the first time,”same
190Speech in loud noise“I haven't had a chance yet to tell you what a jolly little place I think this is.”same
191Speech in loud noise“I'm sure Alexandra hopes you will stay on here, she murmured.”same
192Speech in loud noise“The army found the people in poverty, and left them in comparative wealth.”same
193Speech in loud noise“Our first impressions of people are, in nine cases out of ten, the right impressions.”same
194Speech in loud noise“Dear me, what a chatter box you're getting to be on, remarked the magician, who was pleased with the compliment.”same
195Speech in loud noise“Funchis, who had already approached the door in readiness to leave the road, now recoiled and placed himself once more in the bottom of the live.”same
196Speech in loud noise“It is so new that everywhere we feel the sense of energy.”same
197Speech in loud noise“The attendance was unexpectedly large, and the girls were delighted for seeing great success for their fight.”same
198Speech in loud noise“Would not the loftiest eminences of the city at least be visible?”same
199Speech in loud noise“but the dusk, deepening in the schoolroom, covered over his thoughts. The bell rang,”same
200Speech in loud noise“The mother, in managing the case in this way, relies partly on convincing the reason of the child, and partly on an appeal to her affection.”same
201Common Voice English“The biggest sound stage is leased manually to hold the Golden Eagle Awards.”same
202Common Voice English“The main city on the river is Rostov on Don.”same
203Common Voice English“The requirements are rather high for this job.”same
204Common Voice English“The channel has had critical and popular successes.”same
205Common Voice English“The number still considered retired by the Jet successor team, the Arizona Coyotes.”same
206Common Voice English“He was educated at Falcon.”same
207Common Voice English“A memorial to Charlotte was installed in the Royal Garrison Church at Alderhot.”same
208Common Voice English“Will the weather be freezing in Tennessee?”same
209Common Voice English“The noise came from the base mat.”same
210Common Voice English“While it has long the shelf life, the items will expire in 2026.”same
211Common Voice English“She initially joined the leadership of the Women's Trade Union League.”same
212Common Voice English“QuickRoat always sells Qisp directly to the public through an online store.”same
213Common Voice English“Failed attempts at stopping urination include more petrols and fines.”same
214Common Voice English“The All But One system was composed of several subsystems.”same
215Common Voice English“How should I know anything of her?”same
216Common Voice English“The floorboards looked blue.”same
217Common Voice English“Its capital is the town Pombas.”same
218Common Voice English“However, the multitude of people lived in small cottages or bungalows.”same
219Common Voice English“Dunbar takes action but also takes a bullet, leaving him blind.”same
220Common Voice English“The ground floor has large arcade arches with shop windows for the stores.”same
221Common Voice English“Will it be stormy in Old Fort St. Lucia right now?”same
222Common Voice English“The unincorporated community of Shady Dell is also located in the town.”same
223Common Voice English“We must find out how long to run the experiment.”same
224Common Voice English“Come on, Betty, let's go for a drink and s sink a few.”same
225Common Voice English“Merculus's ambitions are said to have brought him into conflict with Agrippa and others.”same
226Meeting, distant microphone“Or you could do um different like you get with mobile phones different fashions.”same
227Meeting, distant microphone“Togs can some sometimes make out between a thief and a person. So basically these are the three unique features I think belong to you know. Okay, thank you.”same
228Meeting, distant microphone“Can have a side for some preferences, I guess.”same
229Meeting, distant microphone“than like just find it in the program and just say, okay, just how this one does it.”same
230Meeting, distant microphone“because they use them quite frequently in watches.”same
231Meeting, distant microphone“half of the selling price is taken up by building it.”same
232Meeting, distant microphone“My dreadful was the worst yet.”same
233Meeting, distant microphone“Target group. Yes, yes. I don't think we have to I don't think it's a case of worrying about different languages and things like that.”same
234Meeting, distant microphonenothingsame
235Meeting, distant microphone“Um, and I'll be able to not answer them.”same
236Meeting, distant microphone“Is it possible just as an option? When we open it up, people can use their fingers to press the button.”same
237Meeting, distant microphone“Um I think we'll chat about it at the end. Okay.”same
238Meeting, distant microphone“I haven't had some marketing data and basically every product is market driven.”same
239Meeting, distant microphone“But if in a year's time we get feedback from the design fears that show something else is coming in, it's so easy for us to produce that that can be slipped on.”same
240Meeting, distant microphone“Because they're also independent, I'll put that one down again.”same
241Meeting, distant microphone“Um net control is in a way. There is.”same
242Meeting, distant microphone“I have a bit of oh oh you can see it. Okay”same
243Meeting, distant microphone“All it's gonna do is look for separate oh, 'cause then it's gonna go into the NHT search and then turn them into it.”same
244Meeting, distant microphone“I mean sometimes I have methods that don't have any explanation to them at all.”same
245Meeting, distant microphonenothingsame
246Meeting, distant microphone“Good here.”same
247Meeting, distant microphone“Okay, I'm going to look at the functional requirements from a marketing point of view and obviously our starting point with marketing is always market research.”same
248Meeting, distant microphone“Oh no they live there but there's”same
249Meeting, distant microphone“Okay, this is off the top of my head. I was gonna do a bit that too.”same
250Meeting, distant microphone“Just natural. It just seems to be occupied.”same
Interactive. Each clip is a public FSD50K recording with nobody speaking. The dots include the hundred clips of real speech it leaves alone.

This figure is interactive and needs JavaScript.

Figure data for agents: The bouncer, frame by frame (generated from gate.json)

Every clip has real speech in it. Three ways of running the research copy of Phonon-2.5: no clamp, the clamp everywhere, and the clamp only where the bouncer says nobody is talking. Each frame is 80 ms.

A zipper in a pause (10.6 s). Two sentences with a zipper in a 4.5-second pause between them. With no clamp the model types a phantom word over the sound. The clamp everywhere removes it but drops 14 real words. With the bouncer the clamp acts only in the quiet, so the phantom goes and every real word stays.

What was said: “Why, it's in Missouri somewhere on the frontier, I think. We'll get a map. I was afraid it was nearer home.”

SettingPhonon-2.5 writesPhantom wordsReal words changedReal words dropped
No clamp“Why, it's in Missouri somewhere on the frontier, I think. We'll get a map. What? I was afraid it was nearer home.”100
Clamp everywhere“I was afraid it was nearer home.”0014
Clamp with the bouncer“Why, it's in Missouri somewhere on the frontier, I think. We'll get a map. I was afraid it was nearer home.”000

Timeline: speech 0.00 s–4.53 s, pause 4.53 s–9.03 s, sound 6.57 s–7.00 s, speech 9.03 s–10.58 s. The bouncer lets the clamp act on 32 of 133 frames: 5.36 s–7.92 s. The ghost direction at block 18 rises highest at 7.04 s.

Speech: LibriSpeech test-clean 4970-29093-0020 and 4970-29093-0021, CC BY 4.0. Freesound 195070, “Zipper.wav” by punpcklbw, CC BY 3.0, via FSD50K.

A bottle opening in a noisy park (7.3 s). A sentence, 3 seconds of quiet, then a glass bottle being opened. The whole clip sits over the noise of a park. With no clamp the model types a phantom word over the sound. The clamp everywhere removes it but drops 10 real words. With the bouncer the clamp acts only in the quiet, so the phantom goes and every real word stays.

What was said: “Come, come, I called, when no one obeyed. A fire.”

SettingPhonon-2.5 writesPhantom wordsReal words changedReal words dropped
No clamp“Come, come, I called, when no one obeyed. A fire. You across the never and never seen”700
Clamp everywherenothing0010
Clamp with the bouncer“Come, come, I called, when no one obeyed. A fire.”000

Timeline: speech 0.00 s–2.73 s, pause 2.73 s–5.73 s, sound 5.73 s–6.96 s. The bouncer lets the clamp act on 50 of 91 frames: 3.28 s–7.28 s. The ghost direction at block 18 rises highest at 2.72 s.

Speech: LibriSpeech test-clean 5142-33396-0025, CC BY 4.0. Freesound 188903, “glass_bottle_open.wav” by plingativator, CC BY 3.0, via FSD50K. Room noise: DEMAND NPARK (CC BY-SA 3.0); this mix is shared under CC BY-SA 3.0.

A zipper before you start (10.8 s). A zipper, 3.5 seconds of quiet, then a sentence. With no clamp the model types a phantom word over the sound. The clamp everywhere removes it but drops 16 real words. With the bouncer the clamp acts only in the quiet, so the phantom goes and every real word stays.

What was said: “Do you know? Lake? Oh, I really can't tell, but he'll soon tire of country life.”

SettingPhonon-2.5 writesPhantom wordsReal words changedReal words dropped
No clamp“What? Do you know? Lake? Oh, I really can't tell, but he'll soon tire of country life.”100
Clamp everywherenothing0016
Clamp with the bouncer“Do you know? Lake? Oh, I really can't tell, but he'll soon tire of country life.”000

Timeline: sound 0.30 s–0.73 s, pause 0.73 s–4.23 s, speech 4.23 s–10.83 s. The bouncer lets the clamp act on 32 of 136 frames: 0.00 s–2.56 s. The ghost direction at block 18 rises highest at 6.80 s.

Speech: LibriSpeech test-clean 5683-32865-0013, CC BY 4.0. Freesound 195070, “Zipper.wav” by punpcklbw, CC BY 3.0, via FSD50K.

A station announcement (5.7 s). A recorded announcement on a London Underground platform, from a collection of sounds that was meant to have nobody talking in it. The clamp everywhere turns down the model's sense that someone is talking and changes 2 real words. The bouncer hears a voice the whole time, never lets the clamp act, and every word stays.

What the announcement says: “Please mind the gap between the train and the platform.”

SettingPhonon-2.5 writesPhantom wordsReal words changedReal words dropped
No clamp“Please mind the gap between the train and the platform.”000
Clamp everywhere“Please remind the gap between the training and the platform.”020
Clamp with the bouncer“Please mind the gap between the train and the platform.”000

Timeline: speech 0.13 s–4.65 s. The bouncer hears a voice the whole time and never lets the clamp act (0 of 71 frames). The ghost direction at block 18 rises highest at 0.16 s.

Freesound 102302, “Circle Line LU Mind The Gap.mp3” by karthics4, CC0 (the public Freesound copy).

Interactive. Three clips we built from public LibriSpeech, Freesound and DEMAND recordings, and one real station announcement. Every transcript is the research copy of Phonon-2.5’s own output.

What it all adds up to

We went looking for the cause of one strange habit and came out with something closer to a map. A model trained only to turn sound into words keeps careful notes about everything around them, including the language you speak, how fast you talk, the room you are in and whether anyone is talking at all. Reading those notes gives you detectors for free, and writing to them gives you fixes, as long as you also learn when to leave them alone, which is the lesson the bouncer taught us. All of this started on a research copy of the model, and the bouncer is the first piece of it to leave the lab. It now ships in Detta 1.1.3, where it cleans 142 of the 150 noise clips that used to get phantom words (95%), leaves 99 of 100 clips of real speech word for word the same, and never blanks a short command the model heard correctly.

We also have a working theory of why some things get a neuron of their own. Every moment of speech is in some language, at some speed and in some room, so these facts are true of everything the model ever hears, they hold steady across a whole recording, and only one of each applies at a time. That is exactly the kind of thing a network can afford to give a dial of its own. Things that change from moment to moment are spread across many neurons instead, and the clearest case is the one behind the ghosts, the sense of whether anyone is talking right now, which lives in a direction rather than in any single neuron. Our other post, Mapping every neuron in Qwen3-8B, finds the same pattern in a language model, where nearly every single-neuron switch sets who is speaking and how, and none sets whether they are telling the truth. It is not about sheer quantity: across 15 languages, the amount of audio a language has in the largest public collection of European speech does not predict how clean its neuron is. What matters is that a property is always present, not how often any one value of it appears.

We will be open-sourcing the framework behind this work, our tools for finding, testing and editing individual neurons, in the near future, so anyone can run these experiments on their own models.

We don't write research papers. Papers now come out faster than anyone can read them, and good work gets lost in the pile, so we publish our research here, in plain language, where anyone can find it. If you would like to cite this work, please cite this page: Fermion Research, “Reading the neurons of a speech model”, 18 November 2026, https://fermionresearch.com/research/reading-the-neurons-of-a-speech-model/.

Notes for the curious

  1. For the curious, the encoder is a stack of 24 conformer blocks and our neurons are the hidden units of their feed-forward layers, each scored by its average activation over a clip. We grade a neuron with AUROC, the chance that a random clip with the property scores higher than a random clip without it, so a coin flip scores 0.5 and a perfect detector 1.0. A neuron counts only when its score on held-out recordings beats the 95th percentile of twenty scrambled-label searches by at least 0.20, and every push we make on the model is compared with thirty random pushes of the same size. Whisper-small's 98.6% is 1,971 of 2,000 clips (95% CI 97.9 to 99.0%), and the tempo result reproduced on an independent rerun (23.7 to 3.1 errors per hundred words). ↩
  2. Each language neuron scores an AUROC of at least 0.95 on FLEURS speakers it never saw. On 786 VoxPopuli parliament clips in 16 languages the neurons score 99.1%, reading the deepest neuron wherever several tie on FLEURS, against 97.2% for the dedicated VoxLingua107 language-identification model, and the gap holds at 95% confidence however the clips are paired (+0.6 to +3.3 points). A reader trained on a whole block of neurons manages only 27% on the same clips, because it memorised the benchmark's recordings. ↩
  3. The bouncer is a speech reader early in the encoder (block 6 in the original, blocks 4 to 10 in the retrained version) and the clamp acts at block 18. The original bouncer's cutoff came from clean and mildly noisy recordings. On fresh test clips the retrained gate removed 87.3% of 433 phantom words (95% CI 83.8 to 90.1%) and created one new phantom on 1,000 clean clips, and the one noisy collection that moved, URGENT 2024, rose by 0.13 points. Letting Silero VAD gate the same clamp removed 85.9% but changed 42 of the 100 ship-check speech clips and raised URGENT error by about 2 points, and no Silero setting matched the retrained bouncer on both counts. WebRTC VAD marks 53 to 66% of noise frames as speech, which rules it out. ↩

More from the lab

All research