---
license: apache-2.0
library_name: acoustic-book-hash
tags:
  - acoustics
  - fingerprinting
  - impulse-response
  - physical-media
  - numpy
  - cpu
pipeline_tag: other
language:
  - en
---

# acoustic-book-hash

**A closed book rings at its own set of frequencies when tapped.
Those frequencies are a physical fingerprint of the book's
construction.**

A stack of paper is a layered resonator.  Tap the cover, and the
stack rings with a set of modes at frequencies `n · c / (2L)`,
where `c` is the effective wave speed through the pressed pages
and `L` is the stack height.  Two copies of the same edition
share both `c` and `L`.  Different editions differ in at least
one.  The list of peak frequencies is the fingerprint.

## The claim in one sentence

Two copies of the same edition produce nearly identical acoustic
fingerprints; different editions do not.  No training required.

## What it produces

A `Fingerprint` containing:

- `peaks` — up to 8 (frequency, amplitude) pairs in the 200–6000
  Hz band
- `book_id` — an optional string tag

And a database interface:

- `BookDatabase` — a collection of reference fingerprints
- `query(fp)` — returns the best match and the similarity score
- `save(path)` / `load(path)` — JSON persistence

## Install

```bash
pip install numpy
```

No other dependencies.  No model weights.  No downloads.

## Usage

### Build a reference database from WAV files

```bash
python acoustic_book_hash.py record-db mylib.json \
    --book EdA-hardcover-2003 a1.wav a2.wav a3.wav \
    --book EdB-paperback-2010 b1.wav b2.wav
```

### Query a new tap

```bash
python acoustic_book_hash.py query mylib.json unknown_tap.wav
```

Output:

```
query peaks: [809.0, 1219.0, 1600.0, 2402.0, 3205.0] Hz

  EdA-hardcover-2003            0.933
  EdC-folio-1897                0.250
  EdE-pulp-1972                 0.375
  ...

best match: EdA-hardcover-2003  (0.933)
```

### Programmatic

```python
from acoustic_book_hash import (
    BookDatabase, make_fingerprint, load_wav, window_tap,
)

db = BookDatabase.load("mylib.json")
data, sr = load_wav("unknown_tap.wav")
data = window_tap(data, sr)
fp = make_fingerprint(data, sr)
best, all_matches = db.query(fp, threshold=0.6)
print(best)
```

## Benchmarks

Synthetic demo.  Five editions simulated from a physical model.
Each edition's paper speed, stack height, and damping time differ.

### Pairwise similarity matrix

| | EdA | EdB | EdC | EdD | EdE |
|---|---|---|---|---|---|
| EdA | 1.000 | 0.133 | 0.625 | 0.250 | 0.375 |
| EdB | 0.133 | 1.000 | 0.533 | 0.400 | 0.533 |
| EdC | 0.625 | 0.533 | 1.000 | 0.250 | 0.250 |
| EdD | 0.250 | 0.400 | 0.250 | 1.000 | 0.375 |
| EdE | 0.375 | 0.533 | 0.250 | 0.375 | 1.000 |

Diagonal is 1.0 by construction.  Off-diagonal varies with
how close two editions' fundamental frequencies are: EdA and
EdC (800 vs 682 Hz) share more peaks than EdA and EdD (800 vs
1000 Hz), even though both pairs are 150–200 Hz apart.  The
fingerprint is sensitive to *harmonics*, not just the fundamental.

### Robustness

| condition | similarity |
|---|---|
| same tap point, re-tap | 1.000 |
| different tap point | 0.875–1.000 |
| + 1% noise | 1.000 |
| + 3% noise | 0.875 |
| + 5% noise | 0.933 |
| + 10% noise | 1.000 |

Robustness to noise in the demo is better than expected because
the smoothing and peak extraction step removes the noise before
the fingerprint is computed.  Real recordings will have room
reflections on top of noise, which are not in the model.

### Database identification

| query | result | similarity |
|---|---|---|
| fresh tap of EdA | EdA-hardcover-2003 | 1.000 |
| fresh tap of EdC | EdC-folio-1897 | 1.000 |
| fresh tap of EdE | EdE-pulp-1972 | 0.875 |
| EdA + 5% noise | EdA-hardcover-2003 | 0.857 |
| unknown book | *(no match)* | 0.000 |

## Self-test

12 checks, all passing on a clean run:

- physical model produces finite output
- fundamental frequency in the sensible range
- spectral peaks extracted in band
- same edition (different taps): similarity > 0.85
- different editions: similarity < 0.4
- same book + 3% noise: similarity > 0.7
- database correctly identifies a fresh tap
- unknown book correctly rejected
- deterministic simulation
- fingerprint serializes to JSON

## Applications

- **Forgery detection in print-on-demand markets.**  A
  counterfeit uses different paper stock and binding; the
  fingerprints differ.  Requires a reference fingerprint of the
  genuine edition.
- **Edition matching for libraries and rare book dealers.**  A
  dealer can identify which of several similar editions a
  physical copy is without opening it.
- **Copy-level tracking.**  Two copies of the same edition have
  similar fingerprints but not identical ones (paper stock
  varies slightly between printings).  Individual copies can be
  distinguished.
- **Physical provenance.**  A book kept in a humid environment
  has a measurably different fingerprint from the same edition
  kept dry.  The fingerprint encodes the book's history.

## Honest limitations

- **Every number in the benchmarks is from synthetic data.**
  The fingerprints in the demo come from a physical model, not
  from real recordings.  Real books will have more complicated
  spectra (uneven paper, folded pages, dust jackets) and the
  peak extraction may not be as clean.
- **Requires a quiet recording environment.**  In a reverberant
  room, the room's reflections arrive within a few milliseconds
  and overlap the book's own response.  A short window on the
  early response (30 ms) mitigates this but does not eliminate
  it.
- **Requires the same tap point.**  Different tap points change
  the phase pattern, which shifts peak amplitudes but not peak
  frequencies.  Frequency-based comparison is invariant to
  this, but if two editions share the same fundamental, the
  amplitude pattern becomes load-bearing and tap-point
  variance matters.
- **Requires calibration.**  The reference fingerprints must be
  measured once per edition.  There is no training phase.
- **Sensitive to humidity.**  Paper speed varies with moisture
  content.  A book measured at 30% RH and again at 70% RH will
  have shifted mode frequencies.  Real deployments need
  calibration or humidity normalization.
- **The `paper_speed` parameter in the demo is a fitting
  constant.**  Real paper speed varies between 60 and 100 m/s
  depending on paper type.  The five demo editions are one
  plausible family, not a measurement.

## Reference

Part of a series of small tools built in one session:

| tool | reads | answers |
|---|---|---|
| `ir-source-localizer` | impulse responses | where is the source? |
| **`acoustic-book-hash`** | **a tap on a book** | **which edition is it?** |

## License

Apache-2.0