Audio Features

Matft.audio computes audio features compatible with librosa, e.g. the log-mel spectrogram for Whisper.

let audio = MfArray(samples, mftype: .Float) // 16kHz mono
let logmel = Matft.audio.whisper_log_mel(Matft.audio.pad_or_trim(audio), n_mels: 80)
// logmel.shape == [80, 3000]

let spec = Matft.audio.stft(audio, n_fft: 400, hop_length: 160, pad_mode: .reflect) // complex, (201, n_frames)
let mel = Matft.audio.melspectrogram(audio, sr: 16000, n_fft: 400, hop_length: 160, n_mels: 80)
let db = Matft.audio.power_to_db(mel)

See NumPy Mapping › FFT and Audio for the list of functions and their Python counterparts.