/evaluate_asr

An utility script for evaluating ASR results.

Primary LanguagePython

Intro

The script transcribes audio files in a given directory, calculates WER and generates an HTML report containing diffs, WER etc.

Setup

  1. It supports only Python3. First install Python3 if not already installed.
  2. Install virtualenv if not already installed. Run python3 -m pip install --user virtualenv
  3. Create a virtual environment for the script. Navigate to the cloned directory and run python3 -m virtualenv --python=python3 venv .
  4. Install Python package dependencies. python3 -m pip install -r requirements.txt Now you are ready to run the script.

How to run

Before running the script, activate the created virtualenv using source venv/bin/activate (assuming you are in the repository directory where you cloned it).

Input params

  • input-dir: The directory where your audio files and their corresponding ground truth transcripts live. Audio files and transcripts must use same filename. For example, audio1.wav file's transcript name must be audio1.txt
  • audio-extension: wav or mp3, default - wav
  • api-endpoint: The API endpoint URL where the ASR API is deployed. default given.

Run python evaluate_asr.py --input-dir <path-to-your-testset-directory> for executing the script.

The script will save the predicted transcript text files in input-dir with <audio_file_name>-predicted.txt.

It will output an HTML file outside of input-dir with WER report and diffs between ground truth transcripts and predicted transcripts.