R/tempo.R

In jaburgoyne/compmus: Support Functions for Computational Musicology

#' @importFrom magrittr %>%
.sample_tempogram <- function(y, f_s, window_size, hop_size, window_function, cyclic, bpms) {
  window <- window_function(window_size)
  if (cyclic) {
    bpm_octaves <- rep(bpms, 5)
    bpms <- bpms %>%
      tcrossprod(2^(-2:2)) %>%
      as.vector()
  } else {
    bpm_octaves <- bpms
  }
  bases <-
    exp(tcrossprod(-2 * pi * 1i * (bpms / 60) / f_s, 0:(window_size - 1)))
  if (length(y) > window_size) {
    starts <- seq(1, length(y) - window_size, by = hop_size)
  } else {
    starts <- 1
  }
  windowed <- matrix(0, window_size, length(starts))
  for (n in 1:length(starts)) {
    windowed[, n] <- window * y[starts[n]:(starts[n] + window_size - 1)]
  }
  (bases %*% windowed) %>%
    abs() %>%
    magrittr::raise_to_power(2) %>%
    as.vector() %>%
    tibble::tibble(
      time = rep((starts + window_size / 2) / f_s, each = length(bpms)),
      bpm = rep(bpm_octaves, times = length(starts)),
      power = .
    ) %>%
    dplyr::group_by(time, bpm) %>%
    dplyr::summarise(power = sum(power)) %>%
    dplyr::mutate(power = power / max(power)) %>%
    dplyr::ungroup()
}

#' Compute a tempogram from Spotify segment onsets
#'
#' Computes a Fourier-based tempogram based on onsets of Spotify segments.
#' Returns a tibble with \code{time}, \code{bpm}, and \code{power} columns.
#' Power is normalised to a max of 1 (Chebyshev norm) within each time point.
#'
#' @param track_analysis Spotify audio analysis as returned by
#'   \code{\link{get_tidy_audio_analysis}}.
#' @param window_size Window size in seconds (default 8).
#' @param hop_size Hop size in seconds (default 1).
#' @param cyclic Boolean stating whether the tempogram should be cyclic (default
#'   not).
#' @param bpms Vector of tempi in beats per minute to include in the tempogram
#'   (default 30--200 for non-cyclic and 80--160 for cyclic, inclusive of all
#'   integer tempi).
#' @param window_function Window function for the Fourier analysis (default
#'   Hamming).
#'
#' @importFrom magrittr %>%
#' @export
#'
#' @examples
#' library(tidyverse)
#' get_tidy_audio_analysis("6PJasPKAzNLSOzxeAH33j2") %>%
#'   tempogram(window_size = 4, hop_size = 2)
tempogram <- function(track_analysis, window_size = 8, hop_size = 1, cyclic = FALSE, bpms = if (cyclic) 80:160 else 30:600, window_function = signal::hamming) {
  if (tibble::is_tibble(track_analysis)) {
    track_analysis <- track_analysis %>% purrr::flatten()
  }

  assertthat::assert_that(
    track_analysis %>% rlang::has_name("duration"),
    track_analysis %>% rlang::has_name("segments"),
    track_analysis %>%
      purrr::pluck("segments") %>%
      rlang::has_name("start"),
    track_analysis %>%
      purrr::pluck("segments") %>%
      rlang::has_name("confidence"),
    msg = "track_analysis is not a Spotify audio analysis."
  )

  onsets <-
    track_analysis %>%
    purrr::pluck("segments") %>%
    dplyr::pull(start) %>%
    magrittr::multiply_by(44100) %>%
    round()
  confidence <-
    track_analysis %>%
    purrr::pluck("segments") %>%
    dplyr::pull(confidence)

  novelty <- rep(0, 44100 * track_analysis[["duration"]])
  novelty[onsets + 1] <- confidence

  .sample_tempogram(
    novelty,
    f_s = 44100,
    window_size = 44100 * window_size,
    hop_size = 44100 * hop_size,
    cyclic = cyclic,
    bpms = bpms,
    window_function = window_function
  )
}