Podcast Transcription LiveBook

Podcast Transcription.livemd

Podcast Transcription

Mix.install([
  {:req, "~> 0.4.14"},
  {:fast_rss, "~> 0.5.0"},
  {:bumblebee, "~> 0.5.3"},
  {:exla, "~> 0.7.1"},
  {:kino, "~> 0.12.3"}
])

Obtain Episodes

rss_feed_url = "https://feeds.fireside.fm/elixiroutlaws/rss"

%{body: rss_body} = Req.get!(rss_feed_url)

{:ok, rss_feed} = FastRSS.parse_rss(rss_body)

# Grab the fields we care about
episodes =
  Enum.map(rss_feed["items"], fn item ->
    %{
      title: item["title"],
      url: item["enclosure"]["url"]
    }
  end)

# For demonstration, limit the number of episodes to download and process
episode_limit = 2

# Establish a temporary directory to store downloaded podcast episodes
download_directory = Path.join(System.tmp_dir!(), "podcast-downloads")
File.mkdir_p!(download_directory)

episodes =
  episodes
  |> Enum.take(episode_limit)
  |> Enum.map(fn episode ->
    filename = URI.parse(episode.url) |> Map.fetch!(:path) |> Path.basename()
    out_path = Path.join(download_directory, filename)

    Req.get!(url: episode.url, into: File.stream!(out_path))

    Map.put(episode, :local_path, out_path)
  end)

Transcribe

# Download and initialize Whisper model
# Note that other models may have higher accuracy at a cost of slower runtime
{: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],
    chunk_num_seconds: 30,
    timestamps: :segments
  )

# Add Homebrew path, necessary for Mac ffmpeg
os_path = System.get_env("PATH")
homebrew_bin_path = "/opt/homebrew/bin"

if :os.type() == {:unix, :darwin} and not String.contains?(os_path, homebrew_bin_path) do
  System.put_env("PATH", os_path <> ":" <> homebrew_bin_path)
end

episodes =
  Enum.map(episodes, fn episode ->
    start_time = DateTime.utc_now()
    transcription_output = Nx.Serving.run(serving, {:file, episode.local_path})
    end_time = DateTime.utc_now()

    Map.merge(episode, %{
      transcription: transcription_output.chunks,
      transcription_processing_seconds: DateTime.diff(end_time, start_time)
    })
  end)

calculate_transcription_speed_ratio = fn episode ->
  audio_length =
    episode.transcription
    |> Enum.map(fn chunk -> chunk.end_timestamp_seconds end)
    |> Enum.max()

  IO.inspect(episode)

  audio_length / episode.transcription_processing_seconds
end

chunk_to_markdown = fn chunk ->
  "- #{chunk.start_timestamp_seconds}: #{chunk.text}"
end

episode_to_markdown = fn episode ->
  speed_ratio = Float.round(calculate_transcription_speed_ratio.(episode), 2)

  """
  # #{episode.title}

  Transcribed by Whisper at #{speed_ratio}x speed.

  ## Transcript

  #{Enum.map(episode.transcription, &chunk_to_markdown.(&1)) |> Enum.join("\n")}
  """
end

Kino.Markdown.new(episode_to_markdown.(Enum.at(episodes, 0)))

Thank you very much @ChristianAlexander

it actually worked !!

do you think using XLA requires more precomplation of the model than using tensorflow ? that's why it takes more time and space ?