Published on: 2023-11-29

By Emer Butler

In our ongoing series on fine-tuning ASR (Automatic Speech Recognition) models, we’re celebrating the development of our Transcribe ASR platform for your transcription and Whisper Model fine-tuning needs. Check it out!

In this last article of our fine-tuning ASR series, we’ll explore the concept of fine tuning ASR models using strategic vocabulary scaling. Grab a coffee, and let’s get started.

Language is a Living Thing

Before we plunge into the reasons why ASR models sometimes stumble over domain-specific terms, let’s demystify the essence of strategic vocabulary scaling.

Imagine you have a robot friend capable of understanding and jotting down everything you say aloud. But, here’s the catch: when you delve into topics like company-specific protocols or the arcane ingredients of your weekend magic spells (yes, we’re curious too), your robotic pal starts scratching its metallic head. Why? Because these words aren’t part of its everyday lingo.

Language is like a living organism, always evolving. Each industry has its unique jargon, making life a tad complicated for ASR models. While these models have made tremendous strides in understanding everyday speech, they tend to response a bit awkwardly when it comes to specialised language.

This model behaviour becomes a real challenge in sectors where precision is paramount, such as the legal world. Imagine using an automatic transcription tool to accurately transcribe technical terms during a crucial court hearing or a critical medical consultation. For an inaccurate transcription in this use-case, well, the consequences could be, well, quite grim

Where does Strategic Vocabulary Scaling come in? 

To bridge this communication gap, we can employ the concept of “strategic vocabulary scaling.” Think of it as handing your robot buddy a secret dictionary, filled with all the peculiar words tied to your company or your unique niche activity.

Tying back to our example earlier, let’s say you’re working with audio data in the legal realm. In this so called judicial space, terms like “affidavit” and “litigation” get thrown around quite casually. With strategic vocabulary scaling, using our legal jargon Dictionary, we can give our robot friend a crash course on these special terms. In other words, we train the model to integrate a special list of words into its vocabulary. This way, it can more accurately identify the words being spoken and can flawlessly transcribe legal discussions. So essentially, we’re training our model (or the proverbial “robot buddy”) to ne a lot smarter, but only in the areas that matter for a specific task.

How can we help you?

As industries grow increasingly specialized, it’s imperative for ASR models to keep pace and accurately transcribe domain-specific jargon. Strategic vocabulary scaling is a method that fine-tunes ASR models by expanding the model’s vocabulary in a specific domains, whether it’s law, medicine, or finance.

Structuring your audio data so that you can make use of techniques like strategic vocabulary scaling is something we specialise in at Enlabeler, and we’d love to help you! Take a look at our Transcribe ASR platform for your transcription and fine-tuning needs! 

By embracing strategic vocabulary scaling techniques, we can turbocharge the accuracy and reliability of ASR models in transcription tasks. This ensures that critical documents, court proceedings, or medical consultations are captured and recorded with precision, benefiting both professionals and clients alike.