Building an AI-enabled Speech Recognition Engine for a Legal-tech Firm

When it comes to speech recognition technologies, the world is still figuring out the best AI strategies to do so with the cleanest possible methods. The Daffodil AI team opted for the transfer learning-based Machine Learning (ML) model to develop an Automated Speech Recognition (ASR) engine for transcribing live court depositions.

The development of the ASR engine and training the machine to listen to audio and transcribe accurately involved a lengthy process. There was plenty of research and innovation that went into the conceptualization of this process as described in the following sections:

Automating Data Preparation

The ASR machine learning model and engine required thousands of hours of data – both audio of depositions and previously transcribed texts. The audio had to be mono-channel, with a 16 kHz sampling rate for the model to be trained to give more accurate outcomes. This preparation of data before being fed to the ASR engine was automated by our team, so that the engine was equipped to decipher several local Nigerian accents and nuances in the vocabulary.

Pre-Training For Data Augmentation

Once the data was fed into the ASR engine, next came the inclusion of optional modules for data augmentation and pre-training wherever required. Pre-training the ASR machine learning model meant fine-tuning its accuracy through transfer learning methodologies. So the audio files were in WAV format of several people with different linguistic nuances saying the same words. Data training was done using a set of 100,000+ words and linguistics samples. The audio had to be converted and the model was monitored for loss and accordingly fixed when needed.

Incremental Readability

The post-processing pipeline of the ASR machine learning model is quite complex in terms of the text produced and may occasionally have readability issues. An Inverse Text Normalization (ITN) mechanism was implemented to take care of delicate changes that improved the overall readability of the transcribed depositions.

Accurate Text Records

Audio of depositions can be transcribed with near-perfect accuracy. The ASR engine has been trained with so much data that the Word Error Rate (WER) and Character Error Rate (CER) of the audio transcription software never rises above 20%. Depositions, conversations, and hearings can be transcribed to the very letter with no redundant elements caused by noise.