diff --git a/test/audio/keep_talking.wav b/test/audio/keep_talking.wav
new file mode 100644
index 0000000..fdea164
Binary files /dev/null and b/test/audio/keep_talking.wav differ
diff --git a/test/audio/looking_to_book_appointment.wav b/test/audio/looking_to_book_appointment.wav
new file mode 100644
index 0000000..0d67b2f
Binary files /dev/null and b/test/audio/looking_to_book_appointment.wav differ
diff --git a/test/audio/nice_to_meet_you.wav b/test/audio/nice_to_meet_you.wav
new file mode 100644
index 0000000..f201ea1
Binary files /dev/null and b/test/audio/nice_to_meet_you.wav differ
diff --git a/test/bench.exs b/test/bench.exs
new file mode 100644
index 0000000..594b44f
--- /dev/null
+++ b/test/bench.exs
@@ -0,0 +1,236 @@
+Mix.install([
+  {:benchee, "~> 1.0"},
+  {:msgpax, "~> 2.0"},
+  {:websockex, "~> 0.4.3"}
+])
+
+defmodule Wave do
+  @moduledoc """
+  This module contains the Wave struct, which defines fields for the WAVE file
+  audio format.
+
+  Similar to hound's WavSpec, which defines:
+  * channels
+  * sample_ate
+  * bits_per_sample
+  * sample_format
+  """
+
+  alias __MODULE__
+
+  @enforce_keys [:audio_data, :format, :sample_rate, :channels]
+
+  @type t :: %__MODULE__{
+          audio_data: binary(),
+          byte_rate: integer()
+        }
+
+  defstruct chunk_id: nil,
+            chunk_size: nil,
+            # FIXME
+            format: "WAVE",
+            subchunk1_id: nil,
+            subchunk1_size: nil,
+            audio_format: nil,
+            channels: 1,
+            sample_rate: nil,
+            byte_rate: nil,
+            block_align: nil,
+            bits_per_sample: nil,
+            subchunk2_id: nil,
+            subchunk2_size: nil,
+            audio_data: nil
+
+  def new(audio_data, format, sample_rate, channels) do
+    %Wave{
+      chunk_id: nil,
+      chunk_size: nil,
+      format: nil,
+      subchunk1_id: nil,
+      subchunk1_size: nil,
+      audio_format: format,
+      channels: channels,
+      sample_rate: sample_rate,
+      byte_rate: nil,
+      block_align: nil,
+      bits_per_sample: nil,
+      subchunk2_id: nil,
+      subchunk2_size: nil,
+      audio_data: audio_data
+    }
+  end
+
+  @doc """
+  Reads a WAVE file and returns binary data with {:ok, data}, or {:error, reason}
+  `data` contains a struct with parsed informiation.
+  """
+  @spec read(String.t()) :: {:ok, Wave.t()} | {:error, String.t()}
+  def read(path) do
+    with true <- String.ends_with?(path, ".wav"),
+         {:ok, data} <- File.read(path) do
+      {:ok, data}
+    else
+      false -> {:error, "This library only accepts WAV files"}
+      {:error, :enoent} -> {:error, "#{path} does not exist"}
+      error -> error
+    end
+  end
+
+  @spec parse([char]) :: map
+  @doc """
+  `parse` returns a map of headers and data for bitstring file data
+  """
+  def parse(file_data) when is_bitstring(file_data) do
+    # TODO: Need to specify the type of the binary data, e.g., big-integer-size(32)?
+    # TODO: Handle RIFX (big-endian encoded) files. Not default, not sure how common this is...
+
+    ### RIFF (Resource Interchange File Format) chunk descriptor ###
+    # Should contain the letters "RIFF" in ASCII form (0x52494646)
+    # 4 bytes - (0-3) - big endian
+    <<
+      chunk_id::32-big,
+      # 4 + (8 + subchunk1_size) + (8 + subchunk2_size) == 36 + subchunk2_size
+      # The size of the rest of the chunk following this number. The size of the
+      # whole file in bytes, minus 8 bytes for the two fields not included in the
+      # count: chunk_id and chunk_size
+      # 4 bytes - (4-7) - little endian
+      chunk_size::32-little,
+      # 4 bytes - (8-11) - big endian
+      # WAVE format contains two sub-chunks: "fmt" and "data"
+      # contains the letters "WAVE" (0x57415645 in big-endian)
+      format::32-big,
+      ### "fmt" subchunk  describing the format of the sound info in the data sub-chunk ###
+      # contains the letters "fmt" - (0x666d7420 big-endian form)
+      # 4 bytes (12-15) - big endian
+      subchunk1_id::32-big,
+      # 16 for PCM. The size of the rest of the subchunk that follows this number
+      # 4 bytes (16-19) - little endian
+      subchunk1_size::32-big,
+      # PCM = 1 (Linear quantization); values other than 1 indicate some form
+      # of compression
+      # 2 bytes (20-22) - little endian
+      audio_format::16-little,
+      # Mono = 1, Stereo = 2, etc.
+      # 2 bytes (22-23) - little endian
+      channels::16-little,
+      # E.g., 8000, 44_100, etc.
+      # 4 bytes (24-27) - little endian
+      sample_rate::32-little,
+      # equal to sample_rate * channels * bits_per_sample / 8
+      # 4 bytes - (28-31) - little endian
+      byte_rate::32-little,
+      # The number of bytes for one sample, including all channels.
+      # equal to channels * bits_per_sample / 8
+      # 2 bytes - (32-33) - little endian
+      block_align::16-little,
+      # 8 bits = 8, 16 bits = 16, etc.
+      # 2 bytes - (34-35) - little endian
+      bits_per_sample::16-little,
+      # TODO: Handle this possibility
+      # extra_param_size field - if PCM, this doesn't exist, otherwise 2 bytes
+      # extra_params - space for extra parameters
+      #
+      ### "data" subchunk - size of the sound info and contains raw sound data ###
+      #
+      # Contains the letters "data" (0x64617461 big endian form)
+      # 4 bytes - (36-39) - big endian
+      subchunk2_id::32-big,
+      # The number of bytes in the data. Can think of as the size of the read
+      # of the subchunk following this number
+      # equal to the num_samples * channels * bits_per_sample / 8
+      # 4 bytes (40-43) - little endian
+      subchunk2_size::32-little,
+      # Actual sound data - little endian
+      audio_data::little-binary
+    >> = file_data
+
+    %Wave{
+      chunk_id: chunk_id,
+      chunk_size: chunk_size,
+      format: format,
+      subchunk1_id: subchunk1_id,
+      subchunk1_size: subchunk1_size,
+      audio_format: audio_format,
+      channels: channels,
+      sample_rate: sample_rate,
+      byte_rate: byte_rate,
+      block_align: block_align,
+      bits_per_sample: bits_per_sample,
+      subchunk2_id: subchunk2_id,
+      subchunk2_size: subchunk2_size,
+      audio_data: audio_data
+    }
+  end
+end
+
+defmodule BenchClient do
+  use WebSockex
+
+  def open_stream(pid) do
+    msg =
+      Msgpax.pack!(%{type: "open", prompt: "Send me responses as fast as possible!"})
+      |> IO.iodata_to_binary()
+
+    WebSockex.send_frame(pid, {:text, msg})
+    pid
+  end
+
+  def close_stream(pid) do
+    msg = Msgpax.pack!(%{type: "close"}) |> IO.iodata_to_binary()
+    WebSockex.send_frame(pid, {:text, msg})
+  end
+
+  def send_audio(pid, chunk) do
+    msg = Msgpax.pack!(%{type: "audio", audio: Msgpax.Bin.new(chunk)})
+    WebSockex.send_frame(pid, {:text, msg})
+  end
+
+  def handle_frame({:text, msg}, state) do
+    unpacked = %{type: type} = Msgpax.unpack!(msg)
+
+    received =
+      case type do
+        "audio" ->
+          %{audio: audio} = unpacked
+          <<token::8, audio::binary>> = audio
+          :audio
+
+        "token" ->
+          :token
+
+        "interrupt" ->
+          :interrupt
+      end
+
+    {:ok, state}
+  end
+end
+
+audio_clips = Path.join(File.cwd!(), "test/audio/") |> File.ls!() |> Enum.map(&Path.expand/1)
+
+{:ok, pid} =
+  WebSockex.start_link("ws://localhost:4001/conversation/websocket/", BenchClient, %{},
+    debug: [:trace]
+  )
+
+Benchee.run(%{
+  "open_close" => fn ->
+    pid = BenchClient.open_stream(pid)
+    BenchClient.close_stream(pid)
+  end,
+  "send_one_audio" => fn ->
+    pid = BenchClient.open_stream(pid)
+    {:ok, data} = Wave.read(audio_clips |> hd)
+
+    Wave.parse(data).audio_data
+    |> :binary.bin_to_list()
+    |> Stream.chunk_every(1000)
+    |> Stream.each(fn chunk ->
+      IO.puts("Sending chunks!")
+      BenchClient.send_audio(pid, chunk)
+    end)
+    |> Stream.run()
+
+    BenchClient.close_stream(pid)
+  end
+})