A noisy voltage across R and L in series: the current keeps the slow part and loses most of the fast one.
RL is also the name on every commit here.

I work on acoustic and speech signal processing, in Taiwan. These are the repositories, by subject.

English · 繁體中文

Kudio
Audio toolkit, speech most of all: I/O, recording, features, noisy-data synthesis, enhancement and the metrics speech papers report, as small functions over a numpy array and a sample rate.
_{docs · PyPI}

KudioStudio
A desktop workbench for kudio. Open a clip, look at it, change it, measure it and build datasets from it, without writing a script first.
_{user guide · PyPI}

KudioEnhance
Trains and runs learned speech-enhancement models on top of kudio: mixing, training and evaluation as three stages of one config file.
_PyPI

SpeechDsp
Speech DSP in pure NumPy and SciPy: STFT, MFCC, VAD, log-MMSE enhancement and metrics that suit imbalanced classes. No librosa and no deep-learning framework.
_PyPI

voice-pathology-detection
Pathological voice detection from sustained vowels: frame-level MFCC with cepstral normalisation, DBN, MLP, SVM and GMM classifiers, a cascade and a flat route, and cross-validation grouped by subject.

voice-spectral-classifier
One-second spectrograms of sustained vowels classified by a small CNN, with MLP and logistic-regression baselines.

voice-clinic-desktop
A research desktop tool for sustained-vowel screening: noise reduction, voice activity detection, MFCC and per-sex models behind a PySide6 window. Not a medical device.

voice-severity-bench
Feature sets for voice severity ratings compared on the same speaker-grouped folds: MFCC front ends, clinical measures and speech-representation layers, with speaker-level bootstrap, ICC and the rater ceiling.

echo-lattice
Room geometry from one microphone: image-source simulation, NLMS impulse-response identification, first-order echo inversion, and a lattice fit for playback latency and source bearing.

micring
One command for the ReSpeaker USB 4-mic array: recording, direction of arrival, VAD, DSP tuning, the LED ring and spectrograms, on Windows, Linux and Raspberry Pi.

SymLoc
Wi-Fi RSSI fingerprint localization that holds up when an access point is tampered with: likelihoods combined through elementary symmetric polynomials, with spoofing detection.

wifi-fingerprint-locator
Wi-Fi CSI fingerprint localization from Intel 5300 captures: amplitude, phase and IQ fingerprints into an MLP regressor, compared across split strategies.

wifi-presence-sensing
Indoor occupancy detection and fingerprint localization from Intel 5300 Wi-Fi CSI, with AGC calibration, uniform resampling and EEMD modal features.

radar-stream-decoder
Decoder for the UART TLV stream of TI mmWave radars: packet assembly that resynchronises, the demo, Vital Signs and people-counting formats, and a PySide6 viewer.

equipment-vision-inspector
Reads substation equipment nameplates from site photos: Faster R-CNN detection, perspective rectification, Retinex illumination correction, a character CNN and matching against the equipment registry.

thesis-template
A neutral master's thesis LaTeX template in Traditional Chinese, with a Python CLI that builds the PDF and audits citations, references, figures and word counts.

SaDAE-matlab: the speaker-aware deep denoising autoencoder for speech enhancement, in MATLAB
SE-Basic: speech enhancement with a neural network, MATLAB for the data and Keras for the model
DFMaker: image differences side by side, with OpenCV and PyQt5
MyJetsonNote: setting up a Jetson AGX Xavier and a Xavier NX
MyRaspberryNote: setting up a Raspberry Pi
KWS_compile_install: building ML-KWS-for-MCU on Windows

Mostly Python on NumPy and SciPy, PyTorch where a model is trained, MATLAB in the older work,
and a Jetson or a Raspberry Pi when it has to run on a device.
Taiwan, UTC+8. Reach me through GitHub.

_{The figures are drawn by draw.py, and every trace in them is computed.}