Approach

Journalists often source audio in support of their reporting, through interviews and through online research. These recordings are typically transcribed and translated manually, representing significant effort, time and cost. To help address this challenge for La Presse, the leading French-language news media in Canada, the UBC CIC developed T2, a tool that automatically transcribes and translates audio files in 71 languages, generating results in minutes. The output can then be edited to allow for better contextualization. All text (original and in translation) is searchable. 

T2 will also increase cost-efficiency by providing an easy and low-cost way to transcribe and translate audio data. Language service providers typically charge upwards of $1.75 per minute for transcription and $2.50 per minute for translation. As T2 is an automated service, it is estimated to cost $28.00 per month to process up to 730 hours of audio.

Supporting Artifacts

Architecture Diagram

Technical Details

The T2 app enables users to upload audio files to be transcribed and then translated to their desired language. Amazon currently supports 37 languages for transcription and 71 languages for translation.

From observation, the length of the audio file and the time that takes to transcribe and translate is directly proportional. The bigger the size, the longer it takes.

The numbers below explain the sequence illustrated in the architecture diagram of the main components:

1. An audio file or video file is uploaded to the S3 bucket
2. When the file is saved into an S3 bucket, an AWS Lambda function saves the initial data into the database and triggers an AWS Step Function process to start processing the uploaded file.
3. The AWS Step function orchestrates the following tasks:

• Checks if the transcription is complete.
• Uploads transcription info to a database, converts transcript to plan text and uploads to S3 and starts translation job.
• Checks if the translation job is complete.
• Uploads translation info to the database.
• The translation is processed by Amazon Comprehend to extract keyphrases for both translation and transcribe results and saved into the database for searching capabilities.

Link to the solution on GitHub: https://github.com/UBC-CIC/interview-multi-language-transcription

Amazon Transcribe compatible audio and video files can be found at: https://docs.aws.amazon.com/transcribe/latest/dg/input.html