Shepard Tone stream generation in Clojure

Question

This is my work to generate an infinite Shepard Tone. It is written in Clojure and works by generating repeating streams of incrementing frequencies, converting those to values on a sine wave and then adjusting the amplitude to fade in and fade our the audio.

Multiple concurrent streams gives the illusion of an ever increasing pitch.

• line-data holds configuration values.
• freq-freq-lazy-seq is the infinite stream of looping frequencies
• freq-to-sine creates a function that converts a frequency to a value on a sine wave.
• peak-volume creates a function that adjusts the amplitude of a point on the wave.

freq-to-sine and peak-volume are not pure, they each hold internal state. peak-volume doesn't have to, and could be replaced by another stream. freq-to-sine perhaps does, as it's internal state is adjusted based on the argument passed in.

I'm not yet well versed in Clojure so I'm posting this here to gain criticism from experienced Clojure developers, particularly with regard to how close it is to the idiomatic style, whether I committed any language faux pas and comments and insights on how I perhaps could have written it differently.

I had to severely drop the sample rate to prevent buffer underruns on my machine so suggestions on performance improvement are welcome. I haven't investigated profiling in Clojure yet so I don't know whether the underruns are caused by something being slow, or just because it's doing a lot.

I'm especially keen to hear verdict on the impure freq-to-sine function and about any alternative pure way to achieve the same.

The code is available on github along with a README. Please run the code and enjoy it, Shepard Tones are pretty awesome. https://github.com/danmidwood/flox

(ns flox.rad
  (:require clojure.algo.generic.math-functions)
  (:require clojure.math.numeric-tower))

(import '(javax.sound.sampled AudioSystem DataLine$Info
                              SourceDataLine
                              AudioFormat AudioFormat$Encoding))

(def line-data
  {:sample-rate (/ 44100 10)
   :audio-format (AudioFormat. (/ 44100 10) 16 1 true false)
   :write-size (/ 44100 10)
   :12th-root (clojure.math.numeric-tower/expt 2 (/ 1 12))})


(defn- frequency
  "Calculate the frequency of a note offset from a base frequency."
  ([offset base]
     (* base (clojure.math.numeric-tower/expt (:12th-root line-data) offset))))

(defn- freq-freq-lazy-seq
  "Create a lazy sequence of incrementing looping frequencies at one note per second."
  ([base] (freq-freq-lazy-seq base base (frequency 12 base)))
  ([base min max]
     (lazy-seq
      (cons base
            (if (>= base max)
              (freq-freq-lazy-seq min min max)
              (freq-freq-lazy-seq (frequency (/ 1 (:sample-rate line-data)) base) min max))))))


(defn- freq-to-sine
  "Create a function for values on a sine wave from frequencies.
 The produced fn contains internal state to track the distance traveled along the sine so it is not suitable for concurrant use."
  []
  (let [angle (atom 0.0)]
    (fn
      [frequency]
     (let [increment (* 2 (. Math PI) (/ frequency (:sample-rate line-data)))
           this-sine (clojure.algo.generic.math-functions/sin @angle)]
       (do
         ;; The increment is divided by four to correct the pitch. Why though?
         (swap! angle + (/ increment 4))
         this-sine)))))

(defn- peak-volume
  "Create a function that will accept a sine value and adjust it by a magnitude depending on how far through the cycle we are. This means that notes are loudest around the median and quiet at the end ranges.
 The produced fn contains internal state to track how far into the cycle it is so it is not suitable for concurrant use."
  [frames offset]
  (let [middle (/ frames 2)
        distance (atom offset)]
    (fn
      [sine]
      (let [distance-traveled (mod @distance frames)
            distance-from-middle (/ (clojure.math.numeric-tower/abs (- middle distance-traveled)) frames)
            volume (- 1 (* 2 distance-from-middle))]
        (do
          (swap! distance inc)
          (* sine (clojure.math.numeric-tower/expt volume 3)))))))

(defn- byte-my-sine
  "Transforms a floating point sine value (-1..1) into a signed byte (-128..127)"
  ([sine]
     (byte (* (. Byte MAX_VALUE) sine))))

(defn- create-line
  []
  (let [line (. AudioSystem getSourceDataLine (:audio-format line-data))]
    (doto line
      (.open (:audio-format line-data))
      (.start))))

(defn- merge-values
  [& values]
  (/ (apply + values) (count values)))

(defn- write-audio
  [line data]
  (.write ^SourceDataLine line (byte-array data) 0 (:write-size line-data)))

(defn- emit-audio
  [line data]
  (do
    (write-audio line (take (:write-size line-data) data)))
  (recur line (drop  (:write-size line-data) data)))


(defn- freqs-to-sines
  "Accepts a list of n lists of frequencies and returns a list of n lists of sine values"
  [& freqs]
  (apply map
         #(map (freq-to-sine) %)
         freqs))

(defn start
  [base height threads]
  (emit-audio
   (create-line)
   (map byte-my-sine
        (apply map merge-values
               (map
                #(map
                  (peak-volume
                   (* height (:sample-rate line-data))
                   (* (/ 1 threads) % (* height (:sample-rate line-data))))
                  (map
                   (freq-to-sine)
                   (freq-freq-lazy-seq
                    (frequency (* (/ 1 threads) % height) base)
                    base
                    (frequency height base))))
                (range threads))))))