100 lines
3.0 KiB
Elixir
100 lines
3.0 KiB
Elixir
defmodule WhisperWeb.AudioChannel do
|
|
@moduledoc """
|
|
Phoenix Channel que gestiona la recepción de audio en tiempo real desde el cliente.
|
|
|
|
El audio se envía en chunks binarios y se acumula en un buffer temporal (`AudioBuffer`).
|
|
Al recibir el evento `"stop_audio"`, todos los chunks se combinan y se guardan como un único archivo WAV final.
|
|
"""
|
|
|
|
use Phoenix.Channel
|
|
require Logger
|
|
alias Phoenix.PubSub
|
|
|
|
def join("audio:lobby", _payload, socket) do
|
|
ref = socket_id(socket)
|
|
Logger.info("Cliente conectado al canal audio:lobby")
|
|
{:ok, socket}
|
|
end
|
|
|
|
@doc """
|
|
Maneja la entrada de un chunk de audio desde el cliente.
|
|
|
|
El mensaje binario contiene un encabezado JSON (con la tasa de muestreo) seguido del audio binario.
|
|
Se almacena el chunk en `AudioBuffer` y también se guarda como archivo WAV parcial.
|
|
"""
|
|
def handle_in("audio_chunk", {:binary, raw_binary}, socket) do
|
|
<<header_len::16, rest::binary>> = raw_binary
|
|
<<header::binary-size(header_len), audio::binary>> = rest
|
|
|
|
%{"sample_rate" => rate} = Jason.decode!(header)
|
|
ref = socket_id(socket)
|
|
|
|
Logger.info("Chunk recibido: #{byte_size(audio)} bytes, sample_rate: #{rate}")
|
|
AudioBuffer.append(ref, {rate, audio})
|
|
chunks = AudioBuffer.get_and_clear(ref)
|
|
|
|
start_total = System.monotonic_time(:millisecond)
|
|
|
|
if chunks != [] do
|
|
|
|
Task.start(fn ->
|
|
[{rate, _} | _] = chunks
|
|
full_audio = Enum.map(chunks, fn {_, bin} -> bin end) |> IO.iodata_to_binary()
|
|
|
|
{wav_time, {:ok, path}} =
|
|
:timer.tc(fn ->
|
|
AudioSaver.save_chunk_as_wav(ref, full_audio, rate, "part")
|
|
end)
|
|
model_start = System.monotonic_time(:millisecond)
|
|
Logger.info("WAV guardado en #{div(wav_time, 1000)} ms")
|
|
|
|
|
|
transcription =
|
|
if path do
|
|
case Nx.Serving.batched_run(Whisper.LargeModel.Serving, {:file, path}) do
|
|
%{chunks: chunks} ->
|
|
chunks
|
|
|> Enum.map(& &1.text)
|
|
|> Enum.join(" ")
|
|
|
|
_ ->
|
|
"Transcripción no disponible"
|
|
end
|
|
else
|
|
"Archivo no disponible"
|
|
end
|
|
|
|
|
|
model_end = System.monotonic_time(:millisecond)
|
|
Logger.info("El modelo procesó en #{model_end - model_start} ms")
|
|
|
|
Logger.info("✅ Transcripción:\n#{transcription}")
|
|
|
|
message = %{"chunks" => [%{"text" => transcription}]}
|
|
PubSub.broadcast(Whisper.PubSub, "transcription", {:transcription, %{
|
|
"received_at" => model_start,
|
|
"text" => transcription
|
|
}})
|
|
File.rm!(path)
|
|
|
|
end_total = System.monotonic_time(:millisecond)
|
|
Logger.info("⏱ Total procesamiento stop_audio: #{end_total - start_total} ms")
|
|
end)
|
|
end
|
|
{:noreply, socket}
|
|
end
|
|
|
|
def handle_in("stop_audio", _payload, socket) do
|
|
Logger.info("🛑 Grabación detenida por cliente")
|
|
|
|
ref = socket_id(socket)
|
|
|
|
|
|
{:noreply, socket}
|
|
end
|
|
|
|
|
|
defp socket_id(socket), do: socket.transport_pid |> :erlang.pid_to_list() |> List.to_string()
|
|
|
|
end
|