Skip to contents

Track fundamental frequency using SwiftF0 (ONNX)

trk_pitch_swiftf0.Rd

Estimates F0 by applying a convolutional neural network directly to a short-time spectrogram computed inside the ONNX graph (SwiftF0; (Nieradzik 2025) ). SwiftF0 targets real-time speed (~130 ms for 5 s of audio on CPU) at accuracy competitive with CREPE. No Python or PyTorch required — inference uses ONNX Runtime.

Usage

trk_pitch_swiftf0(listOfFiles, beginTime = 0, endTime = 0, minF = 75, maxF = 400, confidence_threshold = 0.9, toFile = TRUE, explicitExt = "sf0", outputDirectory = NULL, verbose = TRUE)

Arguments

listOfFiles

Character vector of audio file paths. Any format supported by av is accepted.

beginTime

Numeric. Start of analysis window in seconds. Default 0 (file start).

endTime

Numeric. End of analysis window in seconds. Default 0 (file end).

minF

Numeric. Minimum F0 in Hz to treat as voiced. Default 75 Hz (speech). Must be >= 46.875 Hz (model minimum; G1). For music, use 46.875.

maxF

Numeric. Maximum F0 in Hz to treat as voiced. Default 400 Hz (speech). Must be <= 2093.75 Hz (model maximum; C7). For music, use 2093.75.

confidence_threshold

Numeric (0–1). Frames with model confidence below this value are marked unvoiced. Default 0.9.

toFile

Logical. If TRUE, write SSFF output files and return the count written (invisibly). If FALSE, return an AsspDataObj. Default TRUE.

explicitExt

Character. Output file extension. Default "sf0".

outputDirectory

Character. Directory for output files. NULL (default) writes alongside the input file.

verbose

Logical. Print per-file progress. Default TRUE.

Value

If toFile = FALSE: an AsspDataObj with tracks:

f0

REAL32, Hz, n\_frames × 1. Fundamental frequency; 0 in unvoiced frames.

confidence

REAL32, 0–1, n\_frames × 1. Model voicing confidence.

Frame rate: fixed 62.5 Hz (16 ms hop; not configurable — fixed by the ONNX graph's internal STFT). If toFile = TRUE: integer count of files written, returned invisibly.

Details

ONNX Runtime is installed automatically on first use (~30 MB, cached in the R user directory). The model file (~400 KB) is downloaded from the swift-f0-onnx Hugging Face Hub repo on first use (requires the huggingfaceR package and a network connection) and cached in the R user directory; subsequent calls read the cached copy with no network access.

Pre-processing (fixed by model training): resample to 16 kHz mono, normalise to full-scale float32 in [-1, 1]. Framing, STFT (1024-sample window, 256-sample hop), and pitch/confidence estimation all happen inside the ONNX graph — no manual windowing is needed. Post-processing (applied outside the graph, matching the reference implementation): a frame is voiced when confidence > confidence_threshold AND its pitch estimate falls within [minF, maxF]; unvoiced frames get f0 = 0.

References

Nieradzik L (2025). “SwiftF0: Fast and Accurate Monophonic Pitch Detection.” Swift-F0 is a fast and accurate F0 detector using CNN-based approach on STFT spectrograms, achieving 132ms processing time for 5 seconds of audio on CPU, 2508.18440, https://arxiv.org/abs/2508.18440.

Examples

if (FALSE) { # \dontrun{
trk_pitch_swiftf0(
  system.file("samples", "sustained", "a1.wav", package = "superassp"),
  toFile = FALSE
)
} # }