DSP Study Group - Reading audio data from WAV-files
Exploration from the Scicloj DSP Study Group Second meeting - Nov. 08th 2025 and some follow-up investigation
Welcome! These are notes from our second study group session, where we’re learning digital signal processing together using Clojure. We’re following the excellent book Think DSP by Allen B. Downey (available free online).
Huge thanks to Professor Downey for writing such an accessible and free introduction to DSP, and for sharing with us the work-in-progress notebooks of Think DSP 2.
Along with this study group came the idea to have an online creative coding festival around Clojure in the first months of 2026. In this meeting we spent some time brainstorming on how that might look and what the scope could be. The remaining time of the session we looked into downloading and reading WAV-files in Clojure.
Why WAV Files?
The notebooks in Think DSP 2 work with WAV files loaded from GitHub as a basis for further processing, so we need a way to load these as well. After obtaining the file, we need to get at the audio data it contains.
Simplified WAV Format
First, let’s take a superficial look at what data WAV files contain, before we dive into getting the data. A simple WAV file consists of a header and pure audio data following it. There are several iterations on specifications for the WAV format and the format allows for quite some flexibility in placing different metadata in the file, as well as different encodings.
The WAV (Waveform Audio File Format) file format is a RIFF (Resource Interchange File Format) file which stores data in chunks. Each chunk consists of a tag and data. Lets consider a partial example, which corresponds to the way the WAV file we want to read is arranged:
The header comprises of the tag RIFF, its chunk tagged with the specific format WAVE and a subchunk fmt, which describes the contained audio data. This represents some of the header information in a WAV file with a single, 16-bit mono sound channel and 44.100 samples per second.
As we learned in the first session of the DSP study group:
Sound waves are continuous vibrations in the air. To work with them on a computer, we need to sample them - take measurements at regular intervals. The sample rate tells us how many measurements per second. CD-quality audio uses 44,100 samples per second.
These samples are stored in the WAV files data tagged subchunk. Since this is mono sound, there is one frame with one channel per sample. For multiple channels, each frame consists of all channels and their respective sample.
Libraries We’re Using
- Kindly - Visualization protocol that renders our data as interactive HTML elements (through Clay)
- Kindly - Visualization protocol that renders our data as interactive HTML elements (through Clay)
- dtype-next - Efficient numerical arrays and vectorized operations (like NumPy for Clojure)
- Tablecloth - DataFrame library for data manipulation and transformation
- Tableplot - Declarative plotting library built on Plotly
- javax.sound.sampled - Some classes from the Java standard libraries sound package to read WAV Files.
(require '[scicloj.kindly.v4.kind :as kind]
'[clojure.java.io :as io]
'[tech.v3.datatype.functional :as dfn]
'[tablecloth.api :as tc]
'[scicloj.tableplot.v1.plotly :as plotly])(import '(javax.sound.sampled AudioFileFormat
AudioInputStream
AudioSystem)
'(java.io InputStream)
'(java.nio ByteBuffer
ByteOrder))Downloading a WAV File
(defn copy [uri file]
(with-open [in (io/input-stream uri)
out (io/output-stream file)]
(io/copy in out)))(copy tuning-fork-url tuning-fork-path)nilPlaying a WAV File
Kindly can embed a player with a URL, but the sample is extremely loud (it is a tuning fork struck in front of a microphone), so we don’t embed this player.
(kind/audio {:src tuning-fork-url})Here we use a compressed and loudness normalized version of the original file, so you can safely listen to it.
(kind/audio {:src tuning-fork-file-compressed})Reading Metadata from the WAV File
We define a function to collect some metadata from the file.
(defn audio-format [^InputStream is]
(let [file-format (AudioSystem/getAudioFileFormat is)
format (.getFormat file-format)]
{:is-big-endian? (.isBigEndian format)
:channels (.getChannels format)
:sample-rate (.getSampleRate format)
:sample-size-bits (.getSampleSizeInBits format)
:frame-length (.getFrameLength file-format)
:encoding (str (.getEncoding format))}))(with-open [wav-stream (io/input-stream tuning-fork-path)]
(def wav-format
(audio-format wav-stream)))#'dsp.wav-files/wav-formatwav-format{:is-big-endian? false,
:channels 1,
:sample-rate 44100.0,
:sample-size-bits 16,
:frame-length 1177856,
:encoding "PCM_SIGNED"}:is-big-endian? specifies the byte order of audio data with more than 8 :sample-size-bits. :sample-size-bits is the number of bits comprising a sample. The :frame-length is the total amount of frames contained in the audio data.
We don’t use much of that information for now, but it’ll let us peek at what kind of WAV file we’re working with in the future and we can use the information to extend our function for extracting audio data, which we define next.
Reading Audio Data from the WAV File
The bulk of work here is handled by the AudionInputStream, but since it only reads bytes for us, we have to put these together into the correct datatype for each frame manually. For now we just put the data for 16-bit mono WAV files into a short-array.
(defn audio-data [^InputStream is]
(let [{:keys [frame-length]} (audio-format is)
format (-> (AudioSystem/getAudioFileFormat is)
AudioFileFormat/.getFormat)
^bytes audio-bytes (with-open [ais (AudioInputStream. is format frame-length)]
(AudioInputStream/.readAllBytes ais))
audio-shorts (short-array frame-length)
bb (ByteBuffer/allocate 2)]
(dotimes [i frame-length]
(ByteBuffer/.clear bb)
(.order bb ByteOrder/LITTLE_ENDIAN)
(.put bb ^byte (aget audio-bytes (* 2 i)))
(.put bb ^byte (aget audio-bytes (inc (* 2 i))))
(aset-short audio-shorts i (.getShort bb 0)))
audio-shorts))(with-open [wav-stream (io/input-stream tuning-fork-path)]
(def wav-shorts
(audio-data wav-stream)))#'dsp.wav-files/wav-shortsThe difference between the WAV file bytes and the audio data we read is 44 bytes, which is the size of the default header and container.
(with-open [wav-stream (io/input-stream tuning-fork-path)]
(- (count (.readAllBytes wav-stream))
(* 2 (count wav-shorts))))44Striking the Fork
Now that we have read the data we can reduce its amplitude, so we can listen to it safely.
^kind/audio
{:samples (dfn// wav-shorts 4000000.0)
:sample-rate (:sample-rate wav-format)}In fact, the function audio-data above is quite similar to how Clay writes the audio data to a file for us to listen to in the browser, just the reverse of what we did for reading.
Visualizing Waves
Let’s take a look at the sound of a tuning fork.
(let [{:keys [frame-length sample-rate]} wav-format]
(-> {:time (dfn// (range frame-length)
sample-rate)
:value wav-shorts}
tc/dataset
(plotly/layer-line {:=x :time
:=y :value})))