Japanese Whisper

Mix.install(
  [
    {:kino_bumblebee, "~> 0.4"},
    {:exla, "~> 0.7"}
  ],
  config: [nx: [default_backend: EXLA.Backend]]
)

Section

{:ok, model_info} = 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"})

generation_config =
  Bumblebee.configure(
    generation_config,
    max_new_tokens: 100,
    forced_token_ids: [
      {1, Bumblebee.Tokenizer.token_to_id(tokenizer, "<|ja|>")},
      {2, Bumblebee.Tokenizer.token_to_id(tokenizer, "<|transcribe|>")},
      {3, Bumblebee.Tokenizer.token_to_id(tokenizer, "<|notimestamps|>")}
    ]
  )

serving =
  Bumblebee.Audio.speech_to_text(model_info, featurizer, tokenizer, generation_config,
    compile: [batch_size: 1],
    defn_options: [compiler: EXLA]
  )

audio_input = Kino.Input.audio("Audio", sampling_rate: featurizer.sampling_rate)
form = Kino.Control.form([audio: audio_input], submit: "Run")
frame = Kino.Frame.new()

Kino.listen(form, fn %{data: %{audio: audio}} ->
  if audio do
    Kino.Frame.render(frame, Kino.Text.new("Running..."))

    audio =
      audio.data
      |> Nx.from_binary(:f32)
      |> Nx.reshape({:auto, audio.num_channels})
      |> Nx.mean(axes: [1])

    %{results: [%{text: generated_text}]} = Nx.Serving.run(serving, audio)
    Kino.Frame.render(frame, Kino.Text.new(generated_text))
  end
end)

Kino.Layout.grid([form, frame], boxed: true, gap: 16)