Explaining a model that listens
On Shapley values for speech, and why the hardest part is choosing who gets to play.
"Transcribe the audio."
Give a speech-language model that instruction and a three-second clip, and it will produce text. The question I care about comes before whether the text is right: what drove the answer? The sound itself, or the instruction that told the model what to expect?
Shapley values are the principled way to ask. Lloyd Shapley introduced them in 1953 to split the payoff of a cooperative game among its players, and they are the only rule that satisfies a short list of fairness axioms: the contributions add up to the whole, identical players get identical credit, a player who changes nothing gets nothing, and values for combined games add.1 Replace players with input features and payoff with the model's output, and you have the backbone of most modern feature attribution.
The price of fairness
A player's Shapley value is its average marginal contribution over every coalition of the other players. With n players there are 2n coalitions. For a short text prompt, sampling estimators make that tractable. For audio, the encoder emits around 50 frames per second, so a three-second utterance is already 150 frames.
| PLAYERS | COUNT, 3 s CLIP | COALITIONS |
|---|---|---|
| Audio frames, 50 per second | 150 | 2150 ≈ 1.4 × 1045 |
| Words | ≈ 8 | 28 = 256 |
No amount of clever sampling makes 1045 feel small. The usual move is to sample harder. I think the better move is to ask whether we picked the right game.
If a lion could talk
"If a lion could talk, we could not understand him."Ludwig Wittgenstein, Philosophical Investigations2
An attribution over 150 frames is the lion talking. Each frame is a twenty-millisecond slice of spectrum. Telling a person that frame 87 mattered is precise and useless, because nobody hears frame 87. People hear words.
That is the idea behind mllm-shap.3 The trick: before computing attributions, group the dense audio frames into word-aligned phonetic segments. The game becomes small enough to estimate well, and the answer comes back in units a person can read: this word in the audio, that word in the instruction. On LFM2-Audio-1.5B, on a single consumer GPU, this reached over 90% accuracy at 15% of the sampling budget, with about 43 times fewer model evaluations per sample.
Is that cheating?
The fair objection: grouping frames throws information away, so the result is an approximation of the "real" frame-level values.
I don't think there is a real, player-free Shapley value to approximate. The definition starts with a set of players. Change the set and you have asked a different question, with its own exact answer. Frame-level attribution answers a question about the encoder's internals. Word-level attribution answers the question a user actually asked. Neither is the other's approximation.
The hard part moves elsewhere: to find the words, you need to know where each one starts and stops in the signal. That alignment problem is the subject of our follow-up, SGPA, which uses the spectrogram itself to place the boundaries.4
Every explanation answers a question someone chose. The least we can do is choose one a person would ask.
Open questions
- Words are the right unit for speech. What is the right unit for music, or for a cough?
- Sarcasm lives between the words, in timing and pitch. Can a word-level game see it at all?
- The same word said twice is two players. Should symmetry treat them as one?
- L. S. Shapley, "A Value for n-Person Games," in Contributions to the Theory of Games II, Princeton University Press, 1953.
- L. Wittgenstein, Philosophical Investigations, 1953. Part II, xi (§327 of "Philosophy of Psychology: A Fragment" in the 4th edition).
- J. Muszyński, P. Pozorski, M. Ganzha, "mllm-shap," ACL 2026 System Demonstrations, pp. 387–396. aclanthology.org/2026.acl-demo.38
- P. Pozorski, J. Muszyński, M. Ganzha, "SGPA," arXiv:2603.02250, 2026.