Illustrated portrait of Taishi Nakashima Taishi Nakashima

Assistant Professor
Tokyo Metropolitan University
Department of Computer Science
Ono Laboratory

Search this site

Managing audio signal processing experiments

Fixing the comparison conditions

Record the sampling rate, STFT window and hop lengths, and input ordering for every experiment. For randomized operations, also record the seed, generator and execution environment. A shared seed alone does not guarantee identical results across libraries or machines.

sample_rate = 16000
fft_size = 1024
hop_size = 256
seed = 42

[evaluation]
warmup_frames = 100

Save this configuration as config.toml. These values are examples for the lesson; choose them for the recordings and method being evaluated.

Identifying the inputs

Record both the path and SHA-256 of each audio file to track content changes even when filenames stay the same. Associate microphone ordering, excerpt boundaries, resampling and normalization settings with the processed audio.

Saving an execution record

The following code uses only the Python 3.11+ standard library. It saves the configuration and input hash in a new run directory. Run it with an actual recording, for example python record_run.py config.toml mixture.wav.

import hashlib
import json
import platform
import sys
import tomllib
import uuid
from datetime import datetime, timezone
from pathlib import Path

config_path, audio_path = map(Path, sys.argv[1:])
config_bytes = config_path.read_bytes()
config = tomllib.loads(config_bytes.decode("utf-8"))
with audio_path.open("rb") as stream:
    audio_hash = hashlib.file_digest(stream, "sha256").hexdigest()

run_id = uuid.uuid4().hex
run_dir = Path("runs") / run_id
run_dir.mkdir(parents=True, exist_ok=False)
record = {
    "run_id": run_id,
    "created_at": datetime.now(timezone.utc).isoformat(),
    "config": config,
    "input": {"path": str(audio_path), "sha256": audio_hash},
    "python": platform.python_version(),
}
(run_dir / "config.toml").write_bytes(config_bytes)
(run_dir / "run.json").write_text(
    json.dumps(record, ensure_ascii=False, indent=2) + "\n",
    encoding="utf-8",
)

Keep the code revision, dependency lock file and execution command in the same directory to document the conditions needed for a rerun. Associate subsequent results with this run_id and the input identity.

Aggregating processing times

For audio duration TaudioT _ {\mathrm{audio}} and processing time TprocessT _ {\mathrm{process}}, define the real-time factor (RTF) as

RTF=TprocessTaudio.\mathrm{RTF} = \frac{T _ {\mathrm{process}}}{T _ {\mathrm{audio}}} .

State whether timing includes file loading and STFT, and use the same scope when comparing methods. For online processing, save per-frame processing times and their maximum in addition to the mean RTF. A mean RTF below one can still hide individual frames that miss their deadlines.

The frame processing budget

With hop length HH and sampling rate fsf _ {\mathrm{s}}, frames arrive every Δt=H/fs\Delta t = H / f _ {\mathrm{s}}. The example configuration gives 16 ms. Treat this interval separately from windowing and input/output buffer latency.

References

Search this site

Enter words to search this site.

    Search this site →