SAE Dissection of Speech Emotion Recognition

In one paragraph. AI models that score top marks at guessing emotion from a voice were actually built to transcribe words — so we never quite knew whether they listen to the tone of voice or to the words, or which internal parts do the work. We used a tool called a Sparse Autoencoder (SAE) to open six of these models and read their internal "features" one by one. Using a clever dataset where everyone says the same sentences, we found: (1) the models tidy emotion into clean features early but a probe reads it best deeper; (2) the emotion features are acoustic (tone), not lexical (words), and they're shared across models and work on unseen speakers; and (3) when a model can cheat by reading the words, only the speech-recognition models do — and only in their deep layers; Whisper never cheats.

1. The big question

Modern speech models (HuBERT, WavLM, wav2vec2, w2v-BERT, MERT, Whisper) are excellent at guessing the emotion in a clip. But they were originally trained to turn speech into text, so they carry a lot of language information inside them. That raises a worry:

When one of these models detects "anger," is it hearing the angry tone of voice — or is it secretly reading the angry words? A model that leans on the words will break exactly where we need it most: sarcasm, accents, and languages it wasn't trained on.

The usual way to study this ("probing") tells us where in the model emotion becomes readable, but it treats each layer as a black box. It can't tell us which pieces inside encode emotion, or what they react to. That's the gap we fill.

2. Key ideas in plain words

Sparse Autoencoder (SAE): A tool that takes the tangled internal signal of a model and untangles it into a big set of simple, mostly single-meaning "features." Think of turning a messy ball of wires into a labelled fuse-box, so we can point at the few wires that carry "emotion."
Monosemantic feature: A feature that means one thing. We want emotion features that fire for one emotion (say, anger) and stay quiet otherwise — not features that mix several things together.
CREMA-D (our main dataset): 91 actors say the same 12 sentences in 6 emotions. Because the words never change, anything the model learns about emotion here must come from the voice, not the text. This is our built-in control.
The confound check: The big methodological trick. We only trust a feature as an "emotion feature" if it tells us more about the emotion than about the sentence (which captures the sounds of the words) or the speaker. Without this, the model's "emotion" features are mostly just detecting sounds of letters — the classic trap where an SAE rediscovers phonemes, not feelings.
Decodability vs. disentanglement: "Decodable" = a simple classifier can read emotion out. "Disentangled" = emotion sits in a few clean, separate features. These turn out to be different things, peaking at different depths.
Acoustic vs. lexical: Acoustic = how it's said (pitch, loudness, voice quality). Lexical = what is said (the words). The core question of the whole study.
EMIS: A special test set of synthetic speech where the written emotion and the spoken emotion are deliberately mismatched — the perfect way to catch a model "cheating" with the words.

3. The whole process, step by step

The study runs as a numbered pipeline. Here is what each step does and why, in plain terms. (Each one is a single reproducible command in the code.)

Step 0 — Take stock

We inventoried the model activations already saved on the cluster, to see exactly what we had to work with.

Why: avoid re-computing expensive things; build the loader around what's really on disk.

Step 0.5 — Get the right data

We downloaded CREMA-D (the fixed-sentence corpus) and later EMIS (the mismatched-speech corpus), and re-ran each frozen model to save its internal activations frame by frame (not averaged), because an SAE needs lots of examples.

Why: the previously saved activations were averaged to one vector per clip — far too few to train a big SAE. Per-frame gave us ~940,000 training examples from one layer.

Step 1 — Train the SAE

We trained the SAE on those activations. It learns to rebuild the signal using only a few active features at a time.

Why / check: it must rebuild the signal faithfully before we trust its features. It explains 93% of the signal with almost no wasted ("dead") features.

Step 2 — Find the emotion features (with the confound check)

We scored every feature by how much its firing relates to the emotion, then kept only those that are clean (single-emotion) and beat the confound check (more about emotion than about the sentence or speaker).

Result: 394 candidate features → 61 genuine emotion features. Two-thirds of the candidates were really about the words or the speaker — exactly the trap we wanted to avoid.

Step 3 — Acoustic or lexical? (the sanity check)

For each emotion feature we asked: is its firing better explained by acoustic measurements (pitch, loudness…) or by which sentence was spoken?

Why: on fixed-sentence CREMA-D the answer must be acoustic — and 60 of 61 features came out acoustic. This proves our labelling method works before we trust it elsewhere.

Step 4 — Depth sweep

We repeated the analysis at all 25 layers and compared two curves: how well a probe decodes emotion, and how many clean emotion features exist.

Finding: the two peak at different depths (see Finding 1).

Step 5 — Compare the six models

We matched features across the six encoders to see if they "think" alike.

Finding: they share a common emotion-feature basis; Whisper is not special here (see Finding 2).

Step 6 — Test on unseen speakers

We re-ran the emotion test making sure the people in the test were never in training.

Finding: almost no drop — the features are about the emotion, not the person.

Step 7 — The causal test

We checked whether emotion can be decoded from only the acoustic features vs. only the lexical ones.

Finding: acoustic features are enough (0.55), lexical ones aren't (0.43) — the signal is acoustic (see Finding 3).

Step 8 — The "cheating" test (EMIS)

On the mismatched-speech set, we measured how often each model goes with the words instead of the voice, layer by layer.

Finding: only the speech-recognition models cheat, and only deep down; Whisper never does (see Finding 4).

4. What we found (the four headlines)

1. The model tidies emotion up early, but reads it best later

Clean, single-meaning emotion features appear in the early-middle layers (2–7). Yet a simple classifier scores highest deeper (layer 10). So "easy to read off" and "neatly organised" are two different things — emotion gets organised early, then re-mixed for easy reading later.

2. All six models share the same emotion features — and ignore the speaker

Every model lands on a similar set of acoustic emotion features, and they keep working on speakers never seen in training (almost no accuracy drop). The features are about the emotion, not the person. On these fixed sentences, even Whisper looks like the others — its famous difference is about words, which this dataset can't show.

3. It's the tone of voice, not the words

60 of 61 emotion features are explained by acoustic properties, not by which sentence was spoken. And emotion can be decoded from the acoustic features (0.55) far better than from the lexical ones (0.43, barely above random 0.42).

4. When a model can cheat with words, only some do — and only deep down

On mismatched speech, the speech-recognition models (HuBERT, WavLM, wav2vec2) start trusting the words — but only in their deep layers. Whisper never does: it stays grounded in the voice at every layer. The music model (MERT) and w2v-BERT never cheat either.

Figures

overview — The whole study at a glance (six panels): depth, acoustic-vs-lexical, cross-model overlap, features per model, cross-speaker, and the causal test.

depth — Depth: a probe reads emotion best deep (layer 10, blue), but the clean emotion features peak early (layer 7, red). Decoding is not the same as disentangling.

acoustic vs lexical — Each emotion feature scored on acoustic vs. lexical prediction. Nearly all sit on the acoustic side — they track the voice, not the words.

EMIS by layer — The "use-the-words" shortcut by layer. It appears only in the deep layers of the speech-recognition models; Whisper (red) stays voice-grounded at every layer.

sufficiency — Causal test: emotion decodes well from acoustic features (0.55), poorly from lexical ones (0.43) — about the same as random (0.42).

5. The numbers at a glance

What	Result
How faithfully the SAE rebuilds the model's signal	93% of the variance
Genuine emotion features (after the confound check)	61 (from 394 candidates)
Emotion features that are acoustic, not lexical	60 of 61
Emotion organised earliest / read best at layer	L7 / L10
Decode emotion from acoustic / lexical / random features	0.55 / 0.43 / 0.42
Accuracy drop on speakers never seen in training	≈ 0
Whisper "uses-the-words" score on mismatched speech	−0.11 (resists)
HuBERT / WavLM / wav2vec2 (deep layers)	+0.7 to +0.9 (use the words)

6. Recent literature — and how we're different

We did a careful search of 2025–2026 work. The short version: our exact combination has not been done, but one very recent paper uses the same tool for a different purpose, so we position against it carefully.

Paper	What it does	How we differ
AudioSAE (EACL 2026)	Same SAE recipe on two models; even compares features across models.	Emotion is just one of four generic tasks — no emotion-feature interpretation, confound check, fixed-sentence corpus, or acoustic-vs-lexical split. We do all of those, on six models. Closest neighbour.
Pluth et al. (2026)	SAE on Whisper, but only the final layer; finds phonetic features.	No emotion target, no depth sweep.
Mariotte et al. (2026)	SAEs on audio models, same "beyond probing" framing.	Case study is singing technique, not emotion.
Ma et al. (2025)	Mechanistic study of Whisper for emotion.	Uses probing/CKA — no SAEs at all.

Bottom line: our novel core — confound-controlled emotion features on a fixed-sentence corpus across six models, plus the depth and EMIS-by-depth findings — is, as far as we can tell, un-scooped. (Note: arXiv IDs should be double-checked before we cite them in a submission.)

7. What to say in the presentation (per slide)

Aim for ~10 minutes, one idea per slide. Here is a simple talk track anyone can deliver.

Slide 1 (Title): "We used sparse autoencoders to open up speech-emotion models and see which internal features carry emotion — and whether they listen to tone or words."

Slide 2 (The question): "These models top emotion benchmarks, but they were built to transcribe words. Are they hearing the feeling or reading the words? Probing can't tell us; we needed to look inside."

Slide 3 (Method): "An SAE is a fuse-box for the model. CREMA-D's fixed sentences are our control. The confound check is how we avoid fooling ourselves with phoneme detectors. 394 candidates become 61 real emotion features."

Slide 4 (Depth): "A probe reads emotion best deep (layer 10), but the clean features peak early (layer 7). Decoding is not the same as understanding."

Slide 5 (Across models & speakers): "All six models share the same acoustic emotion features, and they work on people they've never heard. The features are about the emotion, not the person."

Slide 6 (Acoustic not lexical): "60 of 61 features track the voice, not the words. Emotion decodes from acoustic features (0.55) far better than lexical ones (0.43)."

Slide 7 (EMIS): "On mismatched speech, the speech-recognition models start trusting the words — but only deep down. Whisper never does. Say the caveat out loud: we use the raw activations here, not the SAE reconstruction, which flips Whisper by mistake."

Slide 8 (Literature): "One recent paper (AudioSAE) uses the same tool, but not for emotion and without our controls. Our affect-specific, controlled analysis is the new part."

Slide 9 (Limits & future): "We rely on the sufficiency test, not ablation. Next: a real-transcript test on IEMOCAP, a cross-language study, and feature steering."

Slide 10 (Takeaways): "Decoding isn't understanding. The emotion these models use is in the voice, it's shared and speaker-independent, and only some models cheat with words — deep down."

8. Likely questions (and answers)

Why CREMA-D and not IEMOCAP?: CREMA-D's fixed sentences are the control that lets us separate tone from words. IEMOCAP has real, varied transcripts — which is exactly why it's our planned next test.
Is 61 features too few?: It's after a deliberately strict triple filter (clear emotion link + single-meaning + beats both confounds). We want purity, not a big count; the looser set of 394 is mostly phonetics.
Does it depend on the SAE settings?: Reconstruction is stable (93%, almost no dead features) and the qualitative findings are robust. We haven't exhaustively swept every setting — that's listed as future work.
How is this different from AudioSAE?: Same tool, different question: they probe four generic attributes on two models; we do confound-controlled emotion-feature interpretation on six, plus the depth and EMIS-by-depth results.
Why does the EMIS reconstruction caveat matter?: If you run that one test on the SAE's rebuilt signal instead of the raw model signal, Whisper's result flips by mistake. The SAE is great for reading features, but not as a stand-in for the raw signal in that probe.

9. Limitations & what's next

Sufficiency, not ablation. Emotion is spread across thousands of features, so removing a few barely hurts — we rely on the "can it decode from only these?" test.
One layer per model for the cross-model and EMIS arms (not the full 25-layer grid for all six).
Next experiments (already coded, waiting on data): a real-transcript causal test on IEMOCAP; a cross-language study (SpeechCraft) to explain why English↔Chinese transfer collapses; feature steering (turning an emotion feature up to change the prediction), which would make our claims causal rather than correlational.

10. Where everything lives

Code: src/sae/ (library) and src/sae/steps/ (the pipeline, one command per step).
Results: results/SAE/ (tables, plots, metrics) and its summary/ folder.
Write-ups: the 4-page report and slides in presentation/; the full technical brief in presentation/SAE_branch_brief.md; the literature search in docs/sae/RELATED_WORK.md.
Everything is on the SAE branch. This page is self-contained — it can be hosted directly on a GitHub page.

Prepared for the team to read and get presentation-ready. Questions → see the technical brief on the SAE branch.