Xav

Elixir wrapper over FFmpeg for reading audio and video files.

See an interview with FFmpeg enthusiast: https://youtu.be/9kaIXkImCAM

Installation

Make sure you have installed FFMpeg (ver. 4.x - 7.x) development packages on your system (see here for installation one-liners) and add Xav to the list of your dependencies:

def deps do
  [
    {:xav, "~> 0.7.0"},
    # Add Nx if you want to have Xav.Frame.to_nx/1
    {:nx, ">= 0.0.0"}
  ]
end

Usage

Decode

decoder = Xav.Decoder.new(:vp8)
{:ok, %Xav.Frame{} = frame} = Xav.Decoder.decode(decoder, <<"somebinary">>)

Decode with audio resampling

decoder = Xav.Decoder.new(:opus, out_format: :f32, out_sample_rate: 16_000)
{:ok, %Xav.Frame{} = frame} = Xav.Decoder.decode(decoder, <<"somebinary">>)

Read from a file:

r = Xav.Reader.new!("./some_mp4_file.mp4")
{:ok, %Xav.Frame{} = frame} = Xav.Reader.next_frame(r)
tensor = Xav.Frame.to_nx(frame)
Kino.Image.new(tensor)

Read from a camera:

r = Xav.Reader.new!("/dev/video0", device?: true)
{:ok, %Xav.Frame{} = frame} = Xav.Reader.next_frame(r)
tensor = Xav.Frame.to_nx(frame)
Kino.Image.new(tensor)

Speech to text:

{:ok, whisper} = Bumblebee.load_model({:hf, "openai/whisper-tiny"})
{:ok, featurizer} = Bumblebee.load_featurizer({:hf, "openai/whisper-tiny"})
{:ok, tokenizer} = Bumblebee.load_tokenizer({:hf, "openai/whisper-tiny"})
{:ok, generation_config} = Bumblebee.load_generation_config({:hf, "openai/whisper-tiny"})

serving =
  Bumblebee.Audio.speech_to_text_whisper(whisper, featurizer, tokenizer, generation_config,
    defn_options: [compiler: EXLA]
  )

# Read a couple of frames.
# See https://hexdocs.pm/bumblebee/Bumblebee.Audio.WhisperFeaturizer.html for default sampling rate.
frames =
    Xav.Reader.stream!("sample.mp3", read: :audio, out_format: :f32, out_channels: 1, out_sample_rate: 16_000)
    |> Stream.take(200)
    |> Enum.map(fn frame -> Xav.Frame.to_nx(frame) end)

batch = Nx.Batch.concatenate(frames)
batch = Nx.Defn.jit_apply(&Function.identity/1, [batch])
Nx.Serving.run(serving, batch) 

Development

To make clangd aware of the header files used in your project, you can create a compile_commands.json file. clangd uses this file to know the compiler flags, include paths, and other compilation options for each source file.

Install bear

The easiest way to generate compile_commands.json from a Makefile is to use the bear tool. bear is a tool that records the compiler calls during a build and creates the compile_commands.json file.

You can install bear with your package manager:

• macOS: brew install bear
• Ubuntu/Debian: sudo apt install bear
• Fedora: sudo dnf install bear

Generate compile_commands.json

After installing bear, you can run it alongside your make command to capture the necessary information.

bear -- mix compile