diff --git a/crates/composer/Cargo.toml b/crates/composer/Cargo.toml
index 09383d2..d415a28 100644
--- a/crates/composer/Cargo.toml
+++ b/crates/composer/Cargo.toml
@@ -8,5 +8,6 @@ bincode = "1"
 clap = { version = "4.2", features = ["derive"] }
 color-eyre = "0.6"
 composer_api = { path = "../composer_api" }
+cpal = "0.15"
 eyre = "0.6"
 rodio = { version = "0.17", features = ["symphonia-wav"] }
diff --git a/crates/composer/src/jukebox.rs b/crates/composer/src/jukebox.rs
index e77d7ee..2d5e7dc 100644
--- a/crates/composer/src/jukebox.rs
+++ b/crates/composer/src/jukebox.rs
@@ -1,7 +1,8 @@
+use crate::sound::AudioOutput;
 use eyre::{Context, Result};
 use rodio::{
     source::{Buffered, SamplesConverter},
-    Decoder, OutputStreamHandle, Source,
+    Decoder, Source,
 };
 use std::{collections::HashMap, fs::File, io::BufReader, path::Path};
 
@@ -48,15 +49,12 @@ impl Jukebox {
         Ok(Self { samples })
     }
 
-    pub(crate) fn play(&self, output_stream: &OutputStreamHandle, sample: Sample) -> Result<()> {
+    pub(crate) fn play(&self, audio_output: &AudioOutput, sample: Sample) {
         let buffer = self
             .samples
             .get(&sample)
             .expect("programmer error, all possible samples should be loaded");
 
-        output_stream
-            .play_raw(buffer.clone())
-            .with_context(|| format!("playing sample {sample:?}"))?;
-        Ok(())
+        audio_output.play(buffer.clone());
     }
 }
diff --git a/crates/composer/src/main.rs b/crates/composer/src/main.rs
index d362690..158489c 100644
--- a/crates/composer/src/main.rs
+++ b/crates/composer/src/main.rs
@@ -1,16 +1,19 @@
 #![warn(clippy::all, clippy::clone_on_ref_ptr)]
 
-use crate::jukebox::{Jukebox, Sample};
+use crate::{
+    jukebox::{Jukebox, Sample},
+    sound::AudioOutput,
+};
 use clap::Parser;
 use composer_api::{Event, DEFAULT_SERVER_ADDRESS};
 use eyre::{Context, Result};
-use rodio::{OutputStream, OutputStreamHandle};
 use std::{
     net::UdpSocket,
     time::{Duration, Instant},
 };
 
 mod jukebox;
+mod sound;
 
 #[derive(Parser, Debug)]
 #[command(author, version, about, long_about = None)]
@@ -27,12 +30,12 @@ fn main() -> Result<()> {
     let socket = UdpSocket::bind(args.address.as_deref().unwrap_or(DEFAULT_SERVER_ADDRESS))?;
     println!("Listening on {}", socket.local_addr()?);
 
-    let (_stream, stream_handle) = OutputStream::try_default()?;
+    let audio_output = AudioOutput::new()?;
 
     let jukebox = Jukebox::new().context("creating jukebox")?;
     let mut stats = Stats { since: Instant::now(), events: 0, total_bytes: 0 };
     loop {
-        match handle_datagram(&socket, &stream_handle, &jukebox) {
+        match handle_datagram(&socket, &audio_output, &jukebox) {
             Ok(bytes_received) => stats.record_event(bytes_received),
             Err(err) => eprintln!("Could not process datagram. Ignoring and continuing. {:?}", err),
         }
@@ -42,7 +45,7 @@ fn main() -> Result<()> {
 /// Block until next datagram is received and handle it. Returns its size in bytes.
 fn handle_datagram(
     socket: &UdpSocket,
-    output_stream: &OutputStreamHandle,
+    audio_output: &AudioOutput,
     jukebox: &Jukebox,
 ) -> Result<usize> {
     // Size up to max normal network packet size
@@ -59,7 +62,7 @@ fn handle_datagram(
         // TODO(Pablo): Play a sound that scales with the number of reports.
         Event::LogStats(_) => todo!(),
     };
-    jukebox.play(output_stream, sample)?;
+    jukebox.play(audio_output, sample);
 
     Ok(number_of_bytes)
 }
diff --git a/crates/composer/src/sound.rs b/crates/composer/src/sound.rs
new file mode 100644
index 0000000..b863e19
--- /dev/null
+++ b/crates/composer/src/sound.rs
@@ -0,0 +1,83 @@
+use cpal::{
+    traits::{DeviceTrait, HostTrait},
+    SampleFormat,
+};
+use eyre::{bail, eyre, Result};
+use rodio::{dynamic_mixer::DynamicMixerController, Source};
+use std::{
+    sync::{
+        atomic::{AtomicU64, Ordering},
+        Arc,
+    },
+    time::{Duration, SystemTime, UNIX_EPOCH},
+};
+
+pub(crate) struct AudioOutput {
+    mixer_controller: Arc<DynamicMixerController<f32>>,
+    /// UNIX timestamp in nanoseconds, as an atomic variable.
+    last_callback_timestamp_ns: Arc<AtomicU64>,
+    _stream: cpal::Stream,
+}
+
+impl AudioOutput {
+    pub(crate) fn new() -> Result<Self> {
+        let cpal_device = cpal::default_host()
+            .default_output_device()
+            .ok_or_else(|| eyre!("no cpal audio output device found"))?;
+        let supported_config = cpal_device.default_output_config()?;
+        let stream_config = supported_config.config();
+        println!(
+            "Using audio device '{}', supported config {:?}, stream config {:?}.",
+            cpal_device.name()?,
+            supported_config,
+            stream_config,
+        );
+        if supported_config.sample_format() != SampleFormat::F32 {
+            bail!("Only F32 sample format supported for now.");
+        }
+
+        let (mixer_controller, mut mixer) =
+            rodio::dynamic_mixer::mixer::<f32>(stream_config.channels, stream_config.sample_rate.0);
+
+        let last_callback_timestamp_ns = Arc::new(AtomicU64::new(0));
+
+        let last_callback_timestamp_ns_alias = Arc::clone(&last_callback_timestamp_ns);
+        let _stream = cpal_device.build_output_stream::<f32, _, _>(
+            &stream_config,
+            move |data_out, _info| {
+                // println!("data_callback: date_out.len(): {}, info: {_info:?}.", data_out.len());
+
+                // TODO(Matej): can we use timestamps in _info instead?
+                // TODO(Matej): specify less strict ordering (in all places for this atomic var)
+                last_callback_timestamp_ns_alias.store(current_timestamp_ns(), Ordering::SeqCst);
+
+                data_out.iter_mut().for_each(|d| *d = mixer.next().unwrap_or(0f32))
+            },
+            |err| eprintln!("Got cpal stream error callback: {err}."),
+            None,
+        )?;
+
+        Ok(Self { mixer_controller, last_callback_timestamp_ns, _stream })
+    }
+
+    pub(crate) fn play<S: Source<Item = f32> + Send + 'static>(&self, source: S) {
+        let ns_since_last_callback = current_timestamp_ns()
+            .saturating_sub(self.last_callback_timestamp_ns.load(Ordering::SeqCst));
+        // println!("ns_since_last_callback: {ns_since_last_callback:>9}.",);
+
+        // We assume that cpal's audio callbacks come at a steady rate, so we delay sounds to play
+        // by the time since last audio callback, so that it gets played at correct offset during
+        // the *next* callback. This is needed to accurately position sounds (sub buffer level).
+        // https://github.com/tonarino/acoustic_profiler/issues/19#issuecomment-1522348735
+        self.mixer_controller.add(source.delay(Duration::from_nanos(ns_since_last_callback)));
+    }
+}
+
+fn current_timestamp_ns() -> u64 {
+    let duration_since_epoch =
+        SystemTime::now().duration_since(UNIX_EPOCH).expect("Unable to get current UNIX time");
+    duration_since_epoch
+        .as_nanos()
+        .try_into()
+        .expect("Cannot convert current UNIX timestamp in nanoseconds into u64")
+}