software-mansion-labs / elixir-ibm-speech-to-text

Elixir client library for IBM Cloud Speech to Text service

IBM Cloud Speech to Text

Elixir client for IBM Cloud Speech to Text service

Installation

The package can be installed by adding :ibm_speech_to_text to your list of dependencies in mix.exs:

def deps do
  [
    {:ibm_speech_to_text, "~> 0.3"}
  ]
end

The docs can be found on hexdocs.pm

Usage

1. Start the client process. For that you need to pass API URL or region as an atom, API key obtained from IBM Cloud console. start_link also accepts parameters for the endpoint, see the docs for more details.

   {:ok, pid} = IBMSpeechToText.Client.start_link(:frankfurt, "API_KEY", model: "en-GB_BroadbandModel")

2. Send "start" event with configuration for speech recognition

   start_message = %IBMSpeechToText.Start{content_type: :flac}
   IBMSpeechToText.Client.send_message(pid, start_message)

3. Start audio streaming

   IBMSpeechToText.Client.send_data(pid, audio_data)

4. Stop streaming by sending "stop" message

   stop_message = %IBMSpeechToText.Stop{}
   IBMSpeechToText.Client.send_message(pid, stop_message)

5. You will receive results via message with struct IBMSpeechToText.Response

   %IBMSpeechToText.Response{
     result_index: 0,
     results: [
       %IBMSpeechToText.RecognitionResult{
         alternatives: [
           %IBMSpeechToText.RecognitionAlternative{
             confidence: 0.87,
             timestamps: nil,
             transcript: "to Sherlock Holmes she's always the woman ",
             word_confidence: nil
           }
         ],
         final: true,
         keywords_result: nil,
         word_alternatives: nil
       }, ...
     ],
     speaker_labels: nil,
     warnings: nil
   }

Testing

Test tagged :external is excluded by default since it contacts the real API and requires an API key provided via config. This can be done by adding config/test.secret.exs file with the following content:

use Mix.Config

config :ibm_speech_to_text, api_key: "YOUR_API_KEY"

Fixture

A recording fragment in test/fixtures comes from an audiobook "The adventures of Sherlock Holmes (version 2)" available on LibriVox

Status

There are a few things that are not implemented in current version:

• parsing "word_alternatives" and "keywords_result" in RecognitionResult
• better way to pass endpoint options to client

Copyright and License

Copyright 2019, Software Mansion

Licensed under the Apache License, Version 2.0