README
U
AnalysisResult
U
AudioBuffer
U
BiquadCoeffs
U
BiquadFilter
U
ChromaNorm
U
Complex
U
ContrastConfig
U
CqtConfig
U
FeatureConfig
U
FeatureMatrix
U
Formant
U
FormantConfig
U
FormantFrame
U
HpssConfig
U
HpssResult
U
LogScale
U
LoudnessResult
U
LpcResult
U
MaskMode
U
MelError
U
MelNorm
U
MelScale
U
MultiChannelAudioBuffer
U
OnsetConfig
U
OnsetMethod
U
PitchTrack
U
StreamingExtractor
U
TempoConfig
U
TempogramResult
U
WindowType
U
YinConfig
V
analyze
V
apply_k_weighting
V
autocorrelation
V
bluestein_fft
V
bluestein_ifft
V
chroma_filterbank
V
chroma_from_spectrogram
V
chroma_stft
V
compute_delta
V
compute_energy
V
compute_rms
V
cqt
V
cqt_frequencies
V
create_streaming_extractor
V
create_window
V
cumulative_mean_normalized_difference
V
dct_ii
V
decode_wav
V
deemphasis
V
dft
V
difference_function
V
durand_kerner
V
fft
V
fft_padded
V
formant_frequencies
V
formant_track
V
formants_frame
V
harmonic_energy_ratio
V
hpss
V
hpss_from_audio
V
idft
V
ifft
V
interchannel_level_difference
V
interchannel_phase_difference
V
istft
V
levinson_durbin
V
log_mel_spectrogram
V
loudness_meter
V
lpc
V
lpc_features
V
lpc_residual
V
lpc_to_lpcc
V
lpcc_features
V
median_filter_1d
V
median_filter_freq
V
median_filter_time
V
mel_filterbank
V
mel_spectrogram
V
mfcc
V
multichannel_mel
V
multichannel_mfcc
V
multichannel_stft
V
onset_detect_audio
V
onset_strength
V
onset_strength_audio
V
pad_reflect
V
peak_pick
V
percussive_energy_ratio
V
phase_vocoder
V
pitch_class_name
V
pitch_shift
V
pitch_yin
V
power_to_db
V
preemphasis
V
resample
V
resample_linear
V
resample_polyphase
V
resample_sinc
V
resample_to_length
V
rfft_padded
V
spectral_centroid
V
spectral_contrast
V
spectral_contrast_audio
V
spectral_flatness
V
spectral_flux
V
spectral_rolloff
V
stft
V
stft_complex
V
super_flux
V
tempogram
V
tempogram_from_onset
V
time_stretch
V
true_peak_db
V
yin_frame
V
zero_crossing_rate

moonmel

Lightweight, portable audio feature extraction engine for MoonBit (Mel Spectrogram, Log-Mel, MFCC)

audio

dsp

mel

mfcc

spectrogram

Download zip

Author

1zhouhaoran1

Version

0.1.0

License

MIT

Repository

github.com/1zhouhaoran1/MoonMel

Last updated

55 minutes ago

Downloads

No dependencies

#Run from source

Install MoonBit with moonc >= 0.10.14. Node.js 22.12+ or 24+ and pnpm 11 are needed only for the browser demo. Source installation does not depend on registry publication.

git clone https://github.com/1zhouhaoran1/MoonMel.git
cd MoonMel
moon run examples/basic --target js

Expected output for the included one-second 440 Hz signal:

STFT: 513 x 63
Mel: 80 x 63
Log-Mel: 80 x 63
MFCC: 13 x 63

On the project's Windows workstation, select the installed isolated compiler first:

. ./tools/dev/activate.ps1

#Browser demonstration

pnpm --dir playground install --frozen-lockfile
pnpm --dir playground run dev

Open http://localhost:5173. predev and prebuild regenerate the browser bindings from current MoonBit sources, so compiler output is not maintained by hand. The page executes generated JavaScript in a Web Worker; WASM is a supported library build target. Audio stays on the device.

Action	Result
Select a synthetic tone, sweep, mixture or silence	See an actual MoonBit spectrogram and waveform
Upload an 8/16-bit PCM WAV	Decode with MoonMel, downmix channels and preview up to ten seconds
Change FFT/hop/Mel/MFCC parameters	Recompute the currently selected audio
Select Log-Mel, STFT or MFCC	Inspect values by pointer or left/right keys on the plot
Play, capture a microphone, export	Hear input, analyze a rolling one-second window, download JSON/CSV

Microphone capture requires HTTPS or localhost and explicit browser permission. It uses AudioWorklet; it is a rolling-window demo, not the incremental streaming library API. The page caps uploaded files at 10 MB and preview analysis at 2,000 frames. Shorten audio or increase hop length when a limit is reached.

#API

let cfg = @moonmel.FeatureConfig::default()
let result = @moonmel.analyze(samples, config=cfg).unwrap()
println(result.mfcc.n_cols)

analyze computes one shared STFT and filterbank, then derives Mel, Log-Mel and MFCC in order. Standalone stft, mel_spectrogram, log_mel_spectrogram, mfcc and the existing feature APIs remain available. WAV input uses @moonmel.decode_wav(bytes); then analyze audio.samples with a config whose sample_rate and f_max match the decoded rate.

All feature matrices use [feature row][time frame]. The complex STFT uses [time frame][frequency bin]. Results use Result[T, MelError]; handle errors for user-controlled input rather than unwrapping. Public raw matrices and low-level kernels retain documented preconditions.

#Numerical evidence and limitations

The machine-readable report is produced by tools/reference/check_parity.py using librosa 1.0.0, NumPy 2.5.3 and SciPy 1.18.1. It compares eight deterministic cases, including short windows, arbitrary/odd FFTs and single-sample input. All Mel/Log-Mel/MFCC cells are compared; STFT checks up to 64 explicitly indexed bins in each frame. The comparison aligns periodic Hann, reflect padding, float64 Slaney filters and orthonormal DCT-II, with atol=rtol=1e-7. Observed maximum error across checked values is about 2.3e-12.

This is not whole-library parity or a competitor speed ranking. In particular, librosa's default STFT padding is constant, whereas MoonMel uses reflect. HTK-style deltas, direct CQT, Chroma, HPSS, loudness and effects have their own documented conventions and are not claimed as drop-in equivalents. The Whisper configuration is a framing preset, not a complete model frontend.

Library	Implementation / fit	Difference relevant to this project
MoonMel	Pure MoonBit; compiled JS/WASM/Native	Native MoonBit integration, inspectable core pipeline; external coverage of advanced features is still limited
librosa	Python / NumPy / SciPy research ecosystem	Reference for this project's aligned core numerical checks
Meyda	JavaScript and Web Audio features	Established browser feature integration
Essentia.js	Essentia C++ through WASM with JS/TS API	Broader music algorithms and model integration

#Checks and continuous integration

node tools/quality/check-toolchain.mjs
node tools/quality/check-architecture.mjs
moon check --deny-warn
moon test --deny-warn
moon test --target js --deny-warn
pnpm --dir playground run build
pnpm --dir playground test

Project CI runs check/build/test and the basic example for WASM, WASM-GC, JS and Native, plus a fresh playground build and binding integration tests. CI validates compiler minimum version. A configured workflow is not evidence that a remote run has passed; consult GitHub Actions for that state.

For Windows native validation, run pwsh -File tools/quality/test-native.ps1; it uses an available MinGW GCC when no compiler override is set and restores its environment afterward. The local MSVC discovery points to missing headers, so it cannot be used as-is.

Optional independent Python validation, using the prepared project environment:

& A:\DevEnv\Envs\moonmel_py\Scripts\python.exe tools/reference/check_parity.py

See Development for a portable uv setup, Acceptance for the nine-item acceptance guide, and Health for reviewed packages and remaining limits. Mooncakes publication is deferred at the owner's request.

moonmel

#Run from source

#Browser demonstration

#API

#Numerical evidence and limitations

#Checks and continuous integration

#License

#AnalysisResult

#AudioBuffer

#BiquadCoeffs

#BiquadFilter

#ChromaNorm

#Complex

#ContrastConfig

#CqtConfig

#FeatureConfig

#FeatureMatrix

#Formant

#FormantConfig

#FormantFrame

#HpssConfig

#HpssResult

#LogScale

#LoudnessResult

#LpcResult

#MaskMode

#MelError

#MelNorm

#MelScale

#MultiChannelAudioBuffer

#OnsetConfig

#OnsetMethod

#PitchTrack

#StreamingExtractor

#TempoConfig

#TempogramResult

#WindowType

#YinConfig

#analyze

#apply_k_weighting

#autocorrelation

#bluestein_fft

#bluestein_ifft

#chroma_filterbank

#chroma_from_spectrogram

#chroma_stft

#compute_delta

#compute_energy

#compute_rms

#cqt

#cqt_frequencies

#create_streaming_extractor

#create_window

#cumulative_mean_normalized_difference

#dct_ii

#decode_wav

#deemphasis

#dft

#difference_function

#durand_kerner

#fft

#fft_padded

#formant_frequencies

#formant_track

#formants_frame

#harmonic_energy_ratio

#hpss

#hpss_from_audio

#idft

#ifft

#interchannel_level_difference

#interchannel_phase_difference

#istft

#levinson_durbin

#log_mel_spectrogram

#loudness_meter

#lpc

#lpc_features

#lpc_residual

#lpc_to_lpcc

AnalysisResult

AudioBuffer

BiquadCoeffs

BiquadFilter

ChromaNorm

Complex

ContrastConfig

CqtConfig

FeatureConfig

FeatureMatrix

Formant

FormantConfig

FormantFrame

HpssConfig

HpssResult

LogScale

LoudnessResult

LpcResult

MaskMode

MelError

MelNorm

MelScale

MultiChannelAudioBuffer

OnsetConfig

OnsetMethod

PitchTrack

StreamingExtractor

TempoConfig

TempogramResult

WindowType

YinConfig

analyze

apply_k_weighting

autocorrelation

bluestein_fft

bluestein_ifft

chroma_filterbank

chroma_from_spectrogram

chroma_stft

compute_delta

compute_energy

compute_rms

cqt

cqt_frequencies

create_streaming_extractor

create_window

cumulative_mean_normalized_difference

dct_ii

decode_wav

deemphasis

dft

difference_function

durand_kerner

fft

fft_padded

formant_frequencies

formant_track

formants_frame

harmonic_energy_ratio

hpss

hpss_from_audio

idft

ifft

interchannel_level_difference

interchannel_phase_difference

istft

levinson_durbin

log_mel_spectrogram

loudness_meter

lpc

lpc_features

lpc_residual

lpc_to_lpcc

lpcc_features

median_filter_1d

median_filter_freq

median_filter_time

mel_filterbank

mel_spectrogram

mfcc