Bootstrapping is a nonparametric approach that might help, if you don't know if or cannot assume that your data are normal.

Draw a sample of size n. Get the k-th percentile via quantile() or similar. Draw another sample of size n, sampling with replacement. Get the k-th percentile. Repeat some B number of times, B being the number of bootstrapped samples.

After B samples, you get a set of B measurements that estimate the k-th percentile. Get the 2.5th and 97.5th percentile of this set of measurements. This the 95% confidence interval around your original k-th percentile measurement.

Example code in R:

bootstrap_sample <- function(x, n, B, p) {
  bstrap <- vector(mode = "numeric", length = B)
  for (i in 1:B) {
    s <- sample(x, n, replace = T)
    bstrap[i] <- quantile(s, p)
  }
  bstrap
}

# x = some vector of signal
# n = length of vector
# B = number of bootstrap samples
# kth = k-th percentile of interest
x <- unlist(read.table(...))
n <- length(x)
B <- 10000
kth <- 0.95

# bootstrap samples
x.95thPercentile_95pctCI_samples <- bootstrap_sample(x, n, B, kth)

# k-th percentile (k = 95%)
x.95thPercentile = quantile(x, kth)

# 95th percentile's 95% confidence interval
x.95thPercentile_95pctCI <- quantile(x.95thPercentile_95pctCI_samples, c(0.025, 0.975))

# bounds (asymmetric)
x.95thPercentile_95pctCI_lower_bound = x.95thPercentile - x.95thPercentile_95pctCI[1]
x.95thPercentile_95pctCI_upper_bound = x.95thPercentile + x.95thPercentile_95pctCI[2]