A speech model forgets your voice halfway through

I ran Anthropic's Jacobian lens on five speech recognisers on a shared-memory 16 GB M1 Pro MacBook. About halfway up, each one stops carrying who is speaking and keeps what was said. Most of the work after that point is spelling, and it is the part that breaks first when you compress the model.

This post was motivated by two colleagues. One of them presented Anthropic's paper on the J-space in a meeting. The J-space is a small set of directions inside a language model that hold the words it is ready to say.1 The other uses Wispr Flow with a microphone standing on a box all day and rarely types. Between the two I started to wonder what the same lens would show inside a model that listens.

I am Persian, so I also wanted a Persian model in the set. That is how I found the work of Reza Sayar (@iamRezaSayar), who trains the Shenava recognisers and releases the models, the data and the evaluation sets in the open.9 In my tests his 115M model beats Whisper large-v3-turbo on Persian at a seventh of the size. It is one of the five models below, and it gave me the part of this post I like most.

One sentence, read out of the model at every second block

Granite Speech hears a Welsh speaker say “Here's one for the joystick generation.” Each cell is what the output head reads at that block. Shaded cells already match the final answer.

What the lens does

A speech recogniser is a stack of blocks. Sound goes in at the bottom and a small output head at the top turns the last block's state into text. The simplest way to look inside is to put that head on a lower block and read what comes out. This is the logit lens,3 and it only works when the lower block already stores things the way the top one does.

The Jacobian lens1, 2 adds one matrix per block: the average effect that a small change at that block has on the top block. You push the state through the matrix first, then through the head. I fitted the matrices with finite differences on 16 to 24 clips per model, and scored every reading on clips the fit had not seen.

ModelSizeWhat I read
Granite Speech 5.0 TurboCTC6470M16 encoder blocks. The output head is also trained at block 8.
Parakeet TDT-CTC7, 8110M17 encoder blocks, one head at the top.
Shenava (Persian)9115MSame design as Parakeet.
Whisper large-v3-turbo10809MThe 4 decoder blocks.
Qwen3-ASR110.6BThe 28 blocks of the language model that writes the text.

The maths, in small steps

Call the state of block ℓ at one position hℓ. It is a list of d numbers (1,024 in Qwen). The head is a table W with one row per token. A token's score is its row times the top state, and the model writes the token with the highest score.

scores = W · norm(htop)

The logit lens puts a lower state into the same formula.

logit lensℓ = W · norm(hℓ)

This assumes that block ℓ keeps its notes in the same coordinates as the top block. Often it does not. The blocks above ℓ are a function F that turns hℓ into htop, and they can rotate and stretch the state on the way.

The Jacobian describes F for small steps. Move the state at block ℓ by a small step δ, and the top state moves by about J times δ. J is a table of d by d numbers. Entry (i, j) says how far number i at the top moves when number j at block ℓ moves by one.

F(hℓ + δ) ≈ F(hℓ) + J δJ = ∂htop / ∂hℓ

J changes with every input, so the lens takes its average over many inputs and positions. That gives one table per block, and the lens is the logit lens with that table in front.

Jℓ = average of ∂htop / ∂hℓ

Jacobian lensℓ = W · norm(Jℓ hℓ)

If Jℓ is the identity, this is the logit lens again. So the Jacobian lens is the logit lens after a translation from the coordinates of block ℓ to those of the top.

The paper computes J by backpropagation on 1,000 text prompts. It averages the effect of each position on itself and on later positions, because a language model only looks back.1, 2 I measured it by hand: add 0.1 to one of the d numbers, run the remaining blocks, and see how far the top moved. That is one column of J, and d such runs give the whole table.

column j of J ≈ ( F(h + ε ej) − F(h) ) / εε = 0.1

A speech encoder looks both ways, so I fitted two versions. One adds the step on every frame at once and the other on a few neighbouring frames. J also shrinks the state, so I scale the result back to the size of a real top state before the head reads it.

The J-space follows from here. Each row of W Jℓ is a direction at block ℓ that belongs to one token, and the lens score of a token is how far the state points along its direction. The paper calls J-space the set of states that are a sum of a few of these directions, 25 at most, with positive weights. It holds under a tenth of the variation in a state, and the paper argues that it is the part the model is ready to say.1 The removal test near the end of this post checks that claim in a speech model.

Two ways to read block ℓ

The logit lens hands the state straight to the head. The Jacobian lens first passes it through Jℓ, a linear stand-in for all the blocks above.
soundblocks 1 to ℓblocks above ℓ head Wtext head Wlogit lens Jhead WJacobian lens state at block ℓ

Text appears in the middle

In all three encoders the first six blocks hold almost nothing the head can read. Granite then jumps from 15% to 86% between blocks 6 and 8. That jump is by design: Granite is trained with its head at block 8 as well as at the top, and it feeds that first guess back into itself. Parakeet and Shenava have no such head, and their text comes in gradually over the second half.

Granite has a readable transcript at block 8. The other two build it slowly.

Share of token-writing frames where the plain reading already gives the final token. Granite's line dips after block 8 as the state leaves the head's layout; the Jacobian lens reads 44% at block 8 and 68% at block 14.

Then the voice goes

I trained a linear classifier on single frames to say whether the speaker is male or female, and tested it on accents, and so on voices, that it had never heard. In Granite it is right 99.6% of the time at block 2 and 64% at block 8. At the top it scores 54%, where a coin gets 50%. Accent goes the same way: 73% at block 4, 33% at block 8.

I expected this to be a side effect of Granite's middle head. Parakeet has no head there and does the same thing two blocks later: 96% at block 8, 65% at block 10, which is also where its text starts to show.

Language is treated differently. A probe for English, German, French or Japanese stays above 98% at every block of Granite. The model drops the speaker and keeps the language, which it still needs in order to write.

Both models lose the speaker at the block where the text arrives

Linear probe on one frame. Gender is tested on unseen accents (chance 50%). Accent has six classes (chance 18%).

Granite, text arrives at block 8

Parakeet, text arrives from block 10

The first draft is spelt by ear

Granite's middle head gives a full transcript at block 8, so I compared it with the final one on 284 clips. The block-8 version reads like someone writing down sounds without knowing the words.

What block 8 wrote, and what the top wrote

Words from the two transcripts of the same clips. The last two rows are cases where the second half made things worse.

Of 4,359 words, 334 change between block 8 and the top. In 55 only the capital letters change, in 69 only the punctuation, and 210 become a different word. The repair is worth having: if you stop at block 8 the word error rate is 9.4%, and the full model gets 6.9%.

Put together, Granite does three jobs in order.

Sixteen blocks, three jobs

Shade shows how often the plain reading of that block matches the final token (measured at every second block).

Persian: four letters for one sound

Persian is an Indo-European language, a distant cousin of English and German. It is written in the Arabic alphabet, which was made for a Semitic language with sounds that Persian does not have. Persian kept the letters and merged the sounds. Four letters are all pronounced z, three are s. The spare letters occur almost only in words borrowed from Arabic, so the sound of a word cannot tell the model which letter to write. It has to know the word.

Same sound, different letter

In the first four rows, every letter after the first comes with Arabic loanwords.
SoundLettersExample
zز ذ ض ظضخیم and زخیم sound the same. Only the first is a word (“thick”).
sس ص ثصد (hundred) and سد (dam) sound the same.
tت ط
hه ح
ghق غ

This makes Persian a clean test of the “spelt by ear” idea, and Shenava shows it. I split its tokens into those that hold a loan-only letter and the rest. At block 8 the plain reading already gives the final token for 21% of the ordinary tokens and for 7% of the loan-letter ones. The gap is still there at block 14 (74% against 64%) and closes at block 16.

Shenava settles the Arabic-loan letters last

Share of frames where the plain reading gives the final token, 80 FLEURS clips. Loan-letter tokens (339 frames) are also rarer, and I did not separate the two effects.

The wrong guesses on the way are often the right sound with the wrong letter. Take the cases where a middle block's token differs from the final one in exactly one letter. At blocks 10 and 12 a fifth of those swaps are between two letters with the same sound. Letter frequency alone would give 0.2%.

The finished text has the same fault. Shenava made 41 single-letter mistakes on these clips, and 18 of them are same-sound swaps. In one clip the speaker says ضخیم and the model writes زخیم. Five times it went the other way and wrote the rarer ص where a plain س was correct.

What shapes these curves is the writing system. The language family leaves no mark that I can find. Granite's lens, fitted on English, carries over to Japanese (78% at block 14) better than to German (53%) or French (51%), which are much closer relatives of English. In Qwen, Persian speech behaves like German and Japanese: the plain reading shows 10 to 15% English-only tokens in the middle blocks, and the Jacobian lens 5% at most.

Qwen showed one more thing on Persian. At block 16 the Jacobian lens gives more Latin-letter tokens than Persian ones (66% against 17%). Many of them describe the token the model has just written. Some spell its sound in Latin letters. Others give its meaning in English or in Chinese. The lens also keeps offering “Tehran”, “Iran” and “Shah” where the clip has nothing of the kind, as if all it knows is that the speech is Persian.

I counted both kinds. For sound, I cut the Persian token and the lens token down to their consonants and asked whether they match. At block 16 they do on 22% of positions, against 1% when the lens token is paired with another position. For meaning, I took 150 Persian words where the lens gave an English-looking or Chinese token, mixed them with 150 control pairs, and labelled the 300 without knowing which was which. 37 of the real pairs were translations and none of the controls. About half of all words give such a token, so roughly one Persian word in eight has its translation as the top reading.

After a Persian word, Qwen's middle blocks hold its sound and its meaning

Examples from the Persian clips. The lens reading is the top token of the Jacobian lens at block 16, at the last token of the word.
The model has just writtenThe lens readsWhat that is
شمالnorthernmeaning, in English
منابعsourcesmeaning, in English
تنشtensionmeaning, in English
زمانtimemeaning, in English
منطقه地区的meaning, in Chinese
باbasound, in Latin letters
شşsound, in Latin letters
لیتونletonsound, in Latin letters

The Latin spelling of a Persian token peaks at block 16

Positions where the lens token has the same consonants as the token just written. 80 clips each: 5,384 Persian tokens, 1,327 Japanese kana. The control pairs lens tokens with other positions (Japanese control under 0.5%).

Persian is not special here. I ran the same two counts on the German and Japanese clips, with the same blind labelling. German words reach their English translation most often. Japanese is lowest, partly because Japanese shares its characters with Chinese, so I counted only English readings there.

In all three languages, Qwen's block 16 holds the English for words it has just written

Block 16, 80 clips per language, 150 words and 150 controls labelled blind. A Japanese “word” is one token.
PersianGermanJapanese
Words with an English or Chinese reading47%35%21%
Of those, a translation25%42%32%
Control pairs labelled a translation0%0%0%
All words whose top reading is the translation12%15%7%
Kana or Persian tokens read as their sound, peak22%no test8%

German has no sound test, because it is already in Latin letters. For Japanese I could only test kana, since a kanji's sound needs a dictionary. Kana tokens are short and often all vowel, so their 8% is a lower bound.

So Qwen does hold meanings in a form that reads as English or Chinese. It holds them for words it has finished. For German and Japanese, the removal test further down says that the next word does not depend on them.

English has Persian's problem for a similar reason: its spelling was fixed centuries ago and its sounds moved on. Granite's draft at block 8 and Shenava's at block 10 are both waiting for the part of the model that knows the word.

Two models, the same stages

Linear CKA13 measures how alike two sets of states are, from 0 to 1, whatever the rotation or scale. On the frames that write a token, Granite's blocks fall into groups. Blocks 2 and 4 are close to each other (0.92). Blocks 10 to 14 are close to each other (0.89 to 0.93). Block 8, the first draft, sits between them.

Parakeet is a quarter of the size and comes from NVIDIA, not IBM. It still lines up with Granite stage for stage: Granite's block 4 matches Parakeet's block 6 at 0.85, and Granite's block 10 matches Parakeet's block 12 at 0.74.

The two models pass through similar states in the same order

Linear CKA on the frames that write a token, from 30 clips. On the other frames the picture is different: Granite's block 12 is unlike every other block there, and I do not know why.

Granite against itself

Granite against Parakeet

Compressed models are wrong inside and right outside

I rounded the weights of each encoder to 8, 6, 4, 3 and 2 bits, with the input layer left alone. Four bits costs nothing you can measure in the transcript.

Four bits is free. Two bits is not.

Word error rate in per cent on 60 British-accent clips (40 Persian clips for Shenava). Darker means further from the full-precision model.

The surprise is how wrong the model is inside while it does this. At 4 bits, Granite's state is off by 9% at block 2 and by 25% at block 12, and the transcript goes from 6.9% to 7.0% errors.

At 4 bits the internal error grows with depth, and the text does not change

Size of the change in each block's state against the full-precision model, relative to the state. Granite's bump at block 8 is the middle head.

At 2 bits Granite writes nothing useful. I then kept one part at a time at full precision to see which part matters. Only one choice helps.

The spelling half is the half that cannot take 2 bits

Granite at 2 bits, with one part kept at full precision. Word error rate on 60 clips; shorter is better.

With the second half intact, the error rate comes back to 16.8%. The state that reaches block 8 in that run is off by 92%, and the second half still recovers five words in six from it. Keeping the first half intact does nothing. If you have to go below 4 bits on a model like this, the second half is where to keep the precision. I tested that on one model only.

It does not go through English

Granite also transcribes German, French and Japanese. The clips for this test come from FLEURS.12 Four bits is safe for all four. At 3 bits English is hurt least and Japanese most.

Three bits costs Japanese far more than English

Character error rate on 80 FLEURS clips per language. At 2 bits every language is near 100%, so that column is left out.

Wendler and colleagues reported in 2024 that some multilingual language models lean on English in their middle layers. I looked for the same thing here. For German speech, I counted how often the lens's best token is one that only occurs in my English transcripts. At block 8 that happens on 0.3% of frames. German-only tokens win on 59%. French and Japanese give the same picture.

German speech is read as German from the first readable block

Share of token-writing frames where the plain reading gives a token found only in the German transcripts, or only in the English ones. Granite, 80 German clips.

One more sign that the languages share their machinery: the lens I fitted on 16 English clips works on the other three. At block 14 it gives the final token on 72% of English frames, 53% of German, 51% of French and 78% of Japanese.

A decoder hides its text until the last block

Whisper is built differently. Its decoder writes one token at a time and has only four blocks. The plain reading finds almost nothing below the top. The Jacobian lens does.

In Whisper only the Jacobian lens can read the lower blocks

Share of tokens where the reading of a decoder block gives the final token. English on 204 clips, Persian on 72.
Plain readingJacobian lens

So the right tool depends on the design. In an encoder trained with CTC14 the plain reading is as good as the Jacobian lens or better until the last few blocks. In Whisper's decoder it is useless and the Jacobian lens is the only one that reads anything.

A language model on top

Qwen3-ASR is a third design. A speech encoder turns the audio into a row of vectors, and a 28-block Qwen3 language model reads them and writes the transcript. I put the lens on the language model.

The text stays hidden for longer here than in any other model in this post. The plain reading gives the final token on 7% of positions at block 16, 42% at block 20 and 72% at block 24. Unlike in Whisper, the plain reading does as well as the Jacobian lens. I first fitted the Jacobians on 16 clips and suspected under-fitting, so I fitted them again on 48. The curve moved by two points at most.

Qwen writes nothing readable until block 20 of 28

Share of transcript positions where the reading of a block gives the final token, on 172 clips.

Qwen drops the speaker too, in a different place. The vectors that enter the language model at the audio positions still give the speaker's gender 98% of the time. Four blocks later that is 77%, and at the top 58%. The positions that write the transcript never receive it: the probe peaks at 61% there. Accent is mostly gone before the language model starts (41%, where Granite's early blocks give 73%).

In Qwen the voice stays at the audio positions and fades there

The same probes on the language model; depth 0 is its input. Transcript positions start at chance (51%), so the probe cannot cheat.

The language test gives a less clean answer than it did for Granite. With the plain reading, the middle blocks read German speech as English-only tokens about as often as German-only ones (19% each at block 20), and German only wins at block 24. With the Jacobian lens the English share stays near 5% or below and German-only tokens lead from block 4. One reading shows some English in the middle, as Wendler's team saw in text models,5 and the other shows almost none. Three times as many fitting clips left this split where it was, so the fit size does not explain it.

Both readings do agree that the language is settled long before the word. For Japanese speech the Jacobian lens reads Chinese or Japanese characters on 43% of positions at block 4, where it gets the right token on none of them. For English speech that share stays under 5%.

German speech in Qwen: only the plain reading sees English

Share of transcript positions where each lens gives a token found only in English transcripts. 80 German clips; the Jacobians were fitted on English.

To choose between them I removed what each lens reads and ran the rest of the model. At every transcript position of the German clips I took 200 dimensions out of the state: the main directions of the English-only tokens, of the German-only tokens, or random ones.

In the space of the plain reading this does nothing special. At block 16, 2% of the written tokens change, the same as for random directions. The model does not use what the plain reading sees there, so its English tokens are no step on the way to the German word. In the space of the Jacobian lens the same removal changes 24% of the tokens, seven times the random figure.

The Jacobian lens reads the part of the state that matters, and it sees almost no English. I take that as the answer: Qwen does not go through English either. One caution remains. In the Jacobian space the English and German token directions share most of their 200 dimensions (78% at block 16), so removing either set hurts equally and this test cannot separate the languages there. With 50 dimensions the German set hurts more, 13% against 8%. The output turned into English on 0.2% of positions at most.

Japanese speech gives the same picture at block 16: 4% of tokens change in the space of the plain reading, as for random directions, and 17% in the space of the Jacobian lens against 6% for random ones. English and Japanese token directions again hurt equally there, although the two languages share no script, so that equality says little about language. Block 20 looked like an exception at first. On the 157 positions where the plain reading gives an English-only token, removing that single direction changes 11% of the written tokens, against 2% for the control. I listed them. All 17 changed tokens are commas: the plain reading says “,” where the model goes on to write the Japanese comma “、”, and without that direction it skips the comma and writes the next word. The 81 positions where the plain reading gives letter fragments such as “s” or “y” lose nothing. So the model carries a comma that is not yet tied to a script. It is punctuation and not English.

Only the Jacobian lens reads directions that Qwen uses

Share of written tokens that change when 200 dimensions are removed at block 16, on 80 German clips (3,306 positions).

Hypotheses that didn't hold up

  • “The model knew the right word.” When Granite writes a wrong word, is the right one visible inside? On 182 wrong words it was in the lens's top five 11% of the time. A random word from another clip was there 11% of the time too.
  • The J++ lens. A later write-up4 improves the Jacobian lens for language models with two changes: it freezes the scale of every layer norm, and it fits eight Jacobians per block and keeps the useful ones. On Parakeet, my version of the second change did nothing and the first one broke the lens.
  • Feedback as the weak point. I guessed that Granite's loop from block 8 back into itself would amplify rounding noise. With the loop switched off the model got worse at full precision, at 4 bits and at 3 bits.

Freezing the layer norms breaks the lens on a speech encoder

Parakeet, 40 unseen clips. Eight expert Jacobians matched the plain Jacobian lens within 3 points (in the table). My J++ follows the write-up with about 20 samples per expert, far fewer than the original.

Limits

All of this ran on one laptop. Each lens is fitted on only 16 to 24 clips. The scores use every clip the fit did not see: about 200 for each English model, 64 for Shenava and 72 for Whisper in Persian. An earlier pass on 11 to 30 clips gave the same numbers to within 2 points. The Jacobians come from finite differences. Quantisation is plain rounding in MLX with groups of 64 weights, with no calibration, so a careful scheme would do better at 3 and 2 bits. The gender labels come from the clip names and have two classes. The clips carry no speaker names, which is why the gender probe is tested on held-out accents.

References

  1. W. Gurnee, N. Sofroniew, A. Pearce, M. Piotrowski, I. Kauvar, R. Chen, A. Soligo, P. Bogdan, E. Ong, R. Wang, T. B. Thompson, D. Abrahams, S. Kantamneni, E. Ameisen, J. Batson and J. Lindsey (2026). Verbalizable Representations Form a Global Workspace in Language Models. Transformer Circuits Thread, Anthropic, 6 July 2026. arXiv:2607.15495.
  2. Anthropic (2026). jacobian-lens, the code for [1].
  3. nostalgebraist (2020). Interpreting GPT: the logit lens. LessWrong.
  4. K. Ayonrinde and J. Lindsey (2026). The J++ Lens. In Circuits Updates, October 2026, with a longer write-up on LessWrong.
  5. C. Wendler, V. Veselovsky, G. Monea and R. West (2024). Do Llamas Work in English? On the Latent Language of Multilingual Transformers. ACL 2024. arXiv:2402.10588.
  6. IBM (2026). Granite Speech 5.0 470M TurboCTC, model card.
  7. NVIDIA (2024). Parakeet TDT-CTC 110M, model card.
  8. D. Rekesh, N. R. Koluguri, S. Kriman and others (2023). Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition. arXiv:2305.05084.
  9. R. Sayar (2026). Shenava 1.0: open streaming Persian ASR and captioning. Hugging Face collection.
  10. A. Radford, J. W. Kim, T. Xu and others (2022). Robust Speech Recognition via Large-Scale Weak Supervision. arXiv:2212.04356.
  11. X. Shi, X. Wang, Z. Guo and others (2026). Qwen3-ASR Technical Report. arXiv:2601.21337.
  12. A. Conneau, M. Ma, S. Khanuja and others (2022). FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech. arXiv:2205.12446.
  13. S. Kornblith, M. Norouzi, H. Lee and G. Hinton (2019). Similarity of Neural Network Representations Revisited. ICML 2019. arXiv:1905.00414.
  14. A. Graves, S. Fernández, F. Gomez and J. Schmidhuber (2006). Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks. ICML 2006.
Models: IBM Granite Speech 5.0 TurboCTC, NVIDIA Parakeet TDT-CTC 110M, Shenava FastConformer 115M, OpenAI Whisper large-v3-turbo, Alibaba Qwen3-ASR 0.6B. Everything ran in MLX with mlx-audio. Clips: 220 British-accent recordings and FLEURS.