Automatic Speech Recognition (ASR) Worker

A component that automatically recognises speech from an audio file and transcribes it into text. The implementation is based entirely on Kiirkirjutaja.

Setup

The worker can be used by running the prebuilt docker image. The container is designed to run in a CPU environment. For a manual setup, please refer to the included Dockerfile and the pip packages specification described in requirements.txt.

The worker depends on the following components:

• RabbitMQ message broker

The following environment variables should be specified when running the container:

• MQ_HOST - RabbitMQ host

• MQ_PORT (optional) - RabbitMQ port (5672 by default)

• MQ_USERNAME - RabbitMQ username

• MQ_PASSWORD - RabbitMQ user password

• MQ_EXCHANGE (optional) - RabbitMQ exchange name (speech-to-text by default)

• MQ_CONNECTION_NAME (optional) - friendly connection name (ASR worker by default)

• MQ_HEARTBEAT (optional) - heartbeat value (600 seconds by default)

• API_HOST - ASR service api endpoint

• API_USERNAME - ASR service API username (user by default)

• API_PASSWORD - ASR service API password (pass by default)

• Optional runtime flags (the COMMAND option):

  • --log-config - path to logging config files (logging/logging.ini by default), logging/debug.ini can be used for debug-level logging
  • --port - port of the healthcheck probes (8000 by default):

• Endpoints for healthcheck probes:

  • /health/startup
  • /health/readiness
  • /health/liveness

Performance and Hardware Requirements

The resource requirements of Kiirkirjutaja apply when running the worker:

• Around 16 GB memory should be enough (probably you can do with less)
• Fairly modern fast CPU (development machine has Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz)
• 4 free CPU cores

Request Format

The worker consumes speech recognition requests from a RabbitMQ message broker and responds with the transcribed text straight to the ASR service.

Requests should be published with the following parameters:

• Exchange name: speech-to-text (exchange type is direct)
• Routing key: speech-to-text.<src> where <src> refers to 2-letter ISO language code of the given text. For example speech-to-text.et
• Message properties:
  • Correlation ID - a UID for each request that can be used to correlate requests and responses.
• JSON-formatted message content with the following keys:
  • correlation_id – same as the message property correlation ID
  • file_extension – the file extension of the uploaded audio file (.wav, .mp3, etc.)

The worker will return a response with the following parameters:

• JSON-formatted message content with the following keys:
  • success – whether the transcription of the audio file was a success or not (boolean).
  • result – the transcribed text.