audio-ear

Language: English | 한국어

Reference/test WAV A/B comparison toolkit for drum and synth sound design.

audio-ear helps you quantify what your ears hear:

• transient click/body balance
• low-end decay mismatch
• pitch center and pitch drift
• resonance/ring/formant shifts
• harmonic balance (including 1.5x partial often tied to OSC2-like character)

Requirements

• Python 3.9+
• ffmpeg, ffprobe, sox in PATH (quick analysis)

Quick Compare

python3 scripts/ffmpeg_sox_quick_compare.py \
  --ref /abs/path/A.wav \
  --test /abs/path/B.wav \
  --json-out /tmp/quick.json \
  --md-out /tmp/quick.md

Detailed Compare

python3 scripts/wav_ear_compare.py \
  --ref /abs/path/A.wav \
  --test /abs/path/B.wav \
  --max-seconds 4.0 \
  --fft-size 2048 \
  --hop-size 512 \
  --json-out /tmp/detail.json \
  --md-out /tmp/detail.md

Preset Effective Compare

python3 scripts/preset_effective_compare.py \
  --a /abs/path/preset_before.json \
  --b /abs/path/preset_after.json \
  --json-out /tmp/preset_effective.json \
  --md-out /tmp/preset_effective.md

One-shot Pipeline

bash scripts/run_ear_pipeline.sh /abs/path/A.wav /abs/path/B.wav /tmp/ear-run \
  --preset-a /abs/path/preset_before.json \
  --preset-b /abs/path/preset_after.json

Using With LLM Agents (Codex/Claude)

You can use audio-ear as an "objective ear" inside an agent loop:

1. compare reference vs current render
2. extract measurable gaps
3. modify synth params or DSP
4. re-render and re-compare

Agent Prompt Template

Use audio-ear to compare:
- ref: /abs/path/ref.wav
- test: /abs/path/current.wav
- preset before: /abs/path/preset_before.json
- preset after: /abs/path/preset_after.json

Tasks:
1) Run scripts/run_ear_pipeline.sh and collect quick/detail/preset_effective reports.
2) Summarize the top 5 actionable mismatches with exact metrics.
3) Propose concrete parameter changes (which field, direction, and expected acoustic effect).
4) Keep tonal balance priority over loudness.

Example: Codex/Claude Iteration Prompt

Compare /tmp/ref_tom_h.wav and /tmp/acidbros_ht_factory.wav using audio-ear.
Then tune HT to match these goals:
- stronger F-like body (1.5x partial) but avoid octave-up dominance
- keep wood/ring character
- reduce dull/dark feel without harsh click

Output:
- a short metric table (before/after)
- exact preset changes you applied
- paths to generated WAV and report files

Example: Stop Conditions For Auto-Tuning

Repeat render+compare up to 10 iterations, stop early when all are true:
- |median_f0_delta_cents| <= 15
- click_to_body_delta_db between -1.0 and +1.0
- |harmonic_h15_to_h1_delta_db| <= 3.0
- air band delta >= -3 dB