Japanese Train Jingle Generation

Final project repo for a music ML pipeline that learns from Japanese train station departure jingles and generates new symbolic and audio examples.

Project Structure

.
|-- src/
|   |-- main.py
|   `-- split_audio.py          # Long-audio -> per-jingle WAV clips
|-- data/
|   |-- raw/                    # Put original long MP3/WAV files here
|   |-- split_audio/            # Generated jingle_001.wav, ...
|   |-- midi/                   # Basic Pitch MIDI transcriptions
|   `-- processed/              # Future features, MIDI, tokens, metadata
|-- requirements.txt
|-- requirements-basic-pitch.txt
`-- README.md

Generated data folders are ignored by git except for .gitkeep placeholders.

Setup

Use Python 3.10+.

pip install -r requirements.txt

The script can use the bundled ffmpeg binary from imageio-ffmpeg, which is in requirements.txt. A system ffmpeg on your PATH also works.

Split A Long Jingle Compilation

Put the source audio in data/raw/, then run:

python src/split_audio.py data/raw/long_jingles.mp3 --output-dir data/split_audio --review-html

This creates files like:

data/split_audio/jingle_001.wav
data/split_audio/jingle_002.wav
data/split_audio/manifest.csv
data/split_audio/review.html

For a first pass on station-jingle compilations, try:

python src/split_audio.py data/raw/long_jingles.mp3 ^
  --output-dir data/split_audio ^
  --min-silence-ms 250 ^
  --relative-silence-db -18 ^
  --keep-silence-ms 150 ^
  --min-clip-ms 1200 ^
  --merge-gap-ms 250 ^
  --review-html ^
  --expected-repeats 2

On macOS/Linux, replace the Windows line-continuation ^ with \.

Tuning Segmentation

Main parameters:

• --min-silence-ms: how long a pause must be to count as a split. Increase this if melodies are split in the middle; decrease it if adjacent jingles are stuck together.
• --silence-thresh-dbfs: absolute threshold, often around -35 to -45 for clean audio. More negative means stricter silence detection.
• --relative-silence-db: automatic threshold based on the input average dBFS. The default uses average volume minus 16 dB.
• --keep-silence-ms: padding kept around each jingle so note tails are not cut off.
• --merge-gap-ms: merges tiny gaps that are probably internal rests.
• --min-clip-ms: removes clicks, announcements, or fragments that are too short to be useful.
• --expected-repeats: adds diagnostics for the Flat score pattern where each jingle plays twice in a row. The manifest and review page mark each pair.

Useful debugging workflow:

python src/split_audio.py data/raw/long_jingles.mp3 --dry-run --verbose
python src/split_audio.py data/raw/long_jingles.mp3 --review-html --overwrite

Open data/split_audio/review.html in a browser to audition every split clip. Use data/split_audio/manifest.csv to find the original timestamps for bad splits.

Next Pipeline Steps

After splitting:

1. Run WAV-to-MIDI transcription with Spotify Basic Pitch on data/split_audio/.
2. Store MIDI/tokenized symbolic data in data/processed/.
3. Train a symbolic model for short melody generation.
4. Train or fine-tune an audio/spectrogram model for continuous generation.
5. Evaluate against simple baselines such as random n-gram symbolic generation and pitch/rhythm distribution matching.

Basic Pitch WAV To MIDI

Spotify Basic Pitch is a Python automatic music transcription tool. Its README currently lists Python 3.7 through 3.11 as compatible, so use a Python 3.11 environment for this step.

On Windows:

py -3.11 -m venv .venv-basic-pitch
.venv-basic-pitch\Scripts\activate
python -m pip install -r requirements-basic-pitch.txt

Then transcribe all split WAVs:

python src/transcribe_basic_pitch.py --input-dir data/split_audio --output-dir data/midi

Useful smoke test:

python src/transcribe_basic_pitch.py --input-dir data/split_audio --output-dir data/midi --limit 3 --overwrite

Outputs:

data/midi/jingle_001.mid
data/midi/jingle_002.mid
data/midi/note_events/jingle_001_notes.csv
data/midi/transcription_summary.csv
data/midi/repeat_pair_summary.csv

For cleaner symbolic training data, inspect transcription_summary.csv and the per-file note-event CSVs. Because each jingle appears twice in a row, also inspect repeat_pair_summary.csv; pairs with status check deserve listening or manual note review. Useful cleanup knobs:

• Increase --onset-threshold to remove extra weak notes.
• Increase --frame-threshold to shorten/remove sustained low-confidence notes.
• Increase --minimum-note-length-ms to remove tiny ornaments or artifacts.
• Use --minimum-frequency and --maximum-frequency to restrict the piano range if Basic Pitch invents very low or very high notes.
• Keep pitch bends out of symbolic training at first. The note-event CSV omits them by default; add --include-pitch-bends only if you decide to model them.