Comparison between Google S2R, Pribor’s Combinatorial Magic and Pribor’s CHE (Contextual Hyper-Embedding)
This document presents the characteristics, divergences and synergies between three approaches: Google S2R, Pribor’s Combinatorial Magic and Pribor’s CHE (Contextual Hyper-Embedding).
1. Google’s S2R *
“S2R” means Speech-to-Retrieval. It is a recent voice search architecture that Google is deploying, which bypasses the explicit speech → text transcription step to try to directly establish a match between the spoken audio and the information sought. The model relies on a dual encoder: one processes the audio, the other the candidate texts, in order to bring their vector representations closer together in the same semantic space.
2. Pribor’s Combinatorial Magic
Combinatorial Magic is a bijective, lossless and fixed-dimensional encoding of simple sentences into 4D or 5D vectors: three symbolic components (Subject, Verb, Object) plus a “meta” register of 8 or 16 bits. It is distinguished by O(1) complexity, total absence of information loss, and perfect interpretability.
3. CHE (Contextual Hyper-Embedding uint8)
CHE is an extremely economical contextual encoding approach, which represents each token by a uint8 integer. Unlike the floating-point attention of Transformers, it avoids matrices and softmax, reducing energy consumption by a factor of up to 5000.
4. Comparative Table
|
Feature |
S2R (Google) |
Combinatorial Magic |
CHE |
|
Data type |
float16 / float32 |
symbolic indices + meta uint8 |
uint8 |
|
Dimension |
512–4096D |
4D / 5D |
1 byte/token |
|
Complexity |
O(n²) |
O(1) |
O(n) linear |
|
Information loss |
with loss |
none |
bounded / quantized |
|
Energy efficiency |
low |
extreme |
extreme (×500–5000) |
|
Interpretability |
low |
total |
medium |
5. Synergies and Integration
The three approaches can be integrated into a hybrid architecture: S2R provides the global semantic geometry, CHE ensures contextual efficiency through uint8 quantisation, and Combinatorial Magic formalises symbolic propositions without loss. This combination gives rise to a family of S2R–CHE–CM models combining semantic generalisation, energy frugality and complete interpretability.
* Ehsan Variani and Michael Riley, Research Scientists, Google Research, « Speech-to-Retrieval (S2R): A new approach to voice search », October 7, 2025