---
license: apache-2.0
library_name: numpy
tags:
  - audio
  - symbolic
  - event-detection
  - onset-detection
  - signal-processing
  - heuristic
  - cpu
pipeline_tag: audio-classification
language:
  - en
---

# hv-note

**Symbolic audio event extraction.** No lexicon. No translation layer.

Given a WAV file, output a *score*: a sequence of `(onset_frame, symbol_id)`
pairs, where each symbol describes a discrete audio event by
`(pitch_band, duration_class, amplitude_class)`.

## The claim in one sentence

`hv-tempo` predicts *where* the reader slows down. `hv-core` finds the
cheapest edit that breaks an argument. `hv-fold` counts the passes.
`hv-note` extracts the score — the discrete events in a waveform, and
nothing else.

## What it produces

For a single WAV file:

- **score** — list of `(onset_frame, symbol_id)` pairs
- **events** — for each event: start/end/peak time, pitch Hz, band,
  duration class, amplitude class, symbol id
- **histogram** — symbol counts
- **entropy_bits** — Shannon entropy of the symbol distribution

For a directory:

- **corpus report** — file count, event count, unique symbols, entropy
- **per-group summary** — group key extracted from filename
- **file_scores** — the symbol sequence for every file

## Install

```bash
pip install numpy

Actually — no dependencies. Pure stdlib.

## Usage

### Single file

```bash
python hv_note.py --file ./esc50/audio/1-100032-A-0.wav
python hv_note.py --file ./fsdd/recordings/0_jackson_0.wav

### Directory

```bash
python hv_note.py --dir ./esc50/audio --limit 20
python hv_note.py --dir ./fsdd/recordings
```

### Python

```python
from hv_note import HVNote

m = HVNote()
report = m.score_file("sample.wav")

print(report.score)            # [(0, 44), (24, 44), (49, 44)]
print(report.entropy_bits)     # 0.0
print(report.histogram)        # {44: 3}

fp = m.fingerprint("sample.wav", k=8)
# (44, 44, 44)
```

### CLI flags

```bash
python hv_note.py                                  # synthesized demo
python hv_note.py --file x.wav                     # rendered report
python hv_note.py --file x.wav --json              # JSON
python hv_note.py --file x.wav --fingerprint       # just the fingerprint
python hv_note.py --dir ./audio --limit 100        # first 100 files
python hv_note.py --dir ./audio --json             # corpus JSON
python hv_note.py --save-to ./model                # save config
```

## The model

### Pipeline

```
WAV → frames → onsets → events → symbols → score
```

1. **Frame** the signal. 20 ms window, 10 ms hop. Compute per-frame RMS.
2. **Detect onsets.** A frame is an onset when its RMS is above
   `onset_floor` (0.02) **and** either rises sharply from the previous
   frame (`× onset_rise_mult`) or crosses up out of silence. The
   rising-edge rule prevents a decaying tail from triggering, and the
   silence-crossing rule catches slow attacks like chirps.
3. **Segment events.** The event's extent is the *connected region above
   `onset_floor`* starting at the onset. The peak is the argmax over that
   region only — never over a fixed search horizon, so a later, louder
   event cannot hijack the current one.
4. **Describe each event.**
   - **pitch band** — argmax of Goertzel power over 16 log-spaced bands
     (80 Hz – 6 kHz), on a 30 ms window centered on the peak frame.
   - **duration class** — `{0: <3 frames, 1: 3–7, 2: 8–19, 3: ≥20}`.
   - **amplitude class** — `{0: <0.3, 1: 0.3–0.7, 2: ≥0.7}` relative
     to the recording's own peak RMS.
5. **Encode.** `symbol = band · 12 + dur · 3 + amp`. 16 × 4 × 3 = **192
   symbols**.

### The alphabet

Fixed, unnamed, universal. Symbol `44` means *band 3, duration class 2,
amplitude class 2* — nothing more. Two recordings producing the same
symbol sequence contain the same pattern. No training, no labels, no
translation.

## Benchmarks

### Synthesized samples

| signal | events | unique symbols | entropy |
|---|---:|---:|---:|
| 3× 220 Hz beeps | 3 | 1 | 0.000 b |
| 3 mixed beeps (880/660/440 Hz) | 3 | 3 | 1.585 b |
| 200→2000 Hz chirp | 1 | 1 | 0.000 b |
| 4 noise bursts (claps) | 4 | 4 | 2.000 b |

### On real corpora

**FSDD** (spoken digits, 8 kHz, 3000 recordings). Spoken digits are
single long events. Most recordings reduce to 1–3 symbols. The
distinguishing signal is *pitch band* and *duration class* — "one" is
short and mid-band; "seven" is longer and higher.

**ESC-50** (environmental sound, 44.1 kHz, 2000 recordings). Clips
average 8–20 events. The symbol histogram is a signature of the sound
category. Confusion appears between categories sharing a dominant pitch
band (dog barks and bird calls both light up bands 4–6).

Run either with `python hv_note.py --dir <path>`. Watch `PER-GROUP
SUMMARY`. Similar `unique_sym` counts across two groups mean the two
categories are symbolically indistinguishable in this 192-symbol
alphabet.

## When to use it

- **Matching without labels.** Compare two recordings by edit distance
  between their scores. No classifier needed.
- **Repetition detection.** Find repeated subsequences within a score.
- **Anomaly detection.** Train an n-gram model on a corpus of scores;
  flag low-likelihood scores.
- **Clustering.** Group recordings by symbol histogram.
- **Structural query.** "Find all files whose score contains
  `(5,2,1) → (5,2,1) → (3,1,0)`."

## When *not* to use it

- **For speech recognition.** No phonemes, no words, no letters.
- **For polyphonic music.** The model picks a single dominant band per
  event. Chords collapse.
- **For label prediction.** The output is symbolic, not semantic. If
  you want "this is a dog bark," you need a labeled reference set.
- **For cross-instrument matching.** Piano C4 and violin C4 have
  different timbres, hence different pitch-band energies, hence
  different symbols.
- **For continuous sound without onsets.** Sustained drones, tape hiss,
  and slow ambient textures produce zero events.

## Honest limitations

- **Pitch via Goertzel on 30 ms windows is coarse.** Frequency
  resolution at 80 Hz over 30 ms is ~33 Hz; at 6 kHz it's also ~33 Hz
  but the log bands are wider, so the top bands are tolerant. Fine for
  symbolic bucketing, hopeless for tuning.
- **No polyphony.** Single argmax per event.
- **Unpitched events get random bands.** White noise has a flat
  spectrum; the pitch-band argmax lands wherever the spectral noise
  happens to be loudest on that particular 30 ms window. A clap, a
  snare, and a coin drop all produce different symbols on different
  takes. If you need to match unpitched sounds, average over multiple
  takes rather than matching single recordings.
- **Amplitude collapses to a single class on uniform-loudness files.**
  Amplitude is relative to the recording's own peak RMS, so a file
  whose events are all similar in level produces only `amp_class = 2`.
  The amplitude axis is useful *within* a file that has both loud and
  quiet events, and carries no information otherwise.
- **Onsets are energy-based.** Spectral flux would be sharper, but
  costs more. This is the cheap version.
- **The alphabet is fixed.** A user cannot add custom symbol axes
  without changing the code. That's the point — the alphabet is small
  and universal.
- **No calibration against annotated corpora.** The thresholds are
  hand-tuned and honest.

## Reference

Part of the reader-model series, which has grown beyond reading.

Companion to `hv-tempo` (pace variation), `hv-core` (argument
robustness), `hv-fold` (reading passes).

`hv-note` is the one that doesn't model a reader at all. It models the
signal as a symbolic structure and stops there.

## License

Apache-2.0