audio

name	audio
title	Audio Perception & Synthesis
description	Hearing for the sandbox. Use for any task with audio reference files (.wav/.mp3/.flac/.ogg/.m4a/.aac/.aiff) or that asks to analyze, transcribe, measure, mix, master, or generate sound. Provides FFT / spectral-peak analysis, mel-spectrogram rendering, waveform sampling/decimation, integrated loudness (LUFS), tempo/BPM detection, resampling, and tone/clip synthesis.
modalities	["audio"]
file_extensions	[".wav",".mp3",".flac",".ogg",".m4a",".aac",".aiff"]
keywords	["audio","sound","music","voice","speech","spectrogram","fft","frequency","spectral","waveform","sample rate","sampling","resample","loudness","lufs","decibel","db","tempo","bpm","beat","pitch","mix","master","mono","stereo","transcribe","podcast"]
requires	["librosa","soundfile","numpy","scipy","pyloudnorm"]
version	1.0.0

Audio Skill — give the sandbox ears

Use this skill whenever the task involves listening to or producing sound. Prefer these helpers over hand-rolled DSP: they wrap librosa, soundfile, scipy, and pyloudnorm with safe defaults.

A typical "audio problem" workflow:

1. info = audio.audio_info(path) — duration, sample rate, channels.
2. peaks = audio.fft_summary(path) — dominant frequencies ("what is in it").
3. audio.spectrogram_png(path, "spectrogram.png") — a visual you can ship.
4. Decide / transform (resample, normalise loudness, mix, synthesise).
5. Save deliverables with audio.save_wav(...).

Toolkit API

from skills import audio

audio.audio_info(path) -> dict
    # {sample_rate, duration_sec, channels, n_samples, frames, format}

audio.load_audio(path, sr=None, mono=True) -> (numpy.ndarray, int)
    # waveform y in [-1, 1], native sample rate unless sr given

audio.fft_summary(path, top_k=8, sr=22050) -> dict
    # rFFT magnitude spectrum -> {dominant_hz:[...], magnitudes:[...],
    #                             spectral_centroid_hz, rms, peak_hz}

audio.sample_waveform(path, n=2000) -> list[float]
    # amplitude envelope decimated to n points (cheap "shape" of the signal)

audio.spectrogram_png(path, out="spectrogram.png", n_mels=128) -> str
    # render a mel-spectrogram PNG; returns the output path

audio.loudness_lufs(path) -> float
    # ITU-R BS.1770 integrated loudness via pyloudnorm

audio.tempo_bpm(path) -> float
    # estimated tempo (beats per minute)

audio.resample(path, target_sr, out) -> str
    # write a resampled copy

audio.synth_tone(freq_hz, seconds, sr=44100, amp=0.5) -> (numpy.ndarray, int)
audio.save_wav(y, sr, out) -> str

Notes:

• fft_summary loads at 22.05 kHz mono by default for speed; pass sr=None to use the file's native rate.
• All functions accept absolute or relative paths (reference files are copied into the working directory).