Speech-to-text engines: The digital way to write documents

In the ancient Hindu religious books, we have read that the Sage Vyasa dictated the Mahabharata to little Ganesha without taking a pause. Ganesha was able to complete the writing of a lengthy epic in a very short span of time. In the current age, we have a speech to text engines to convert the spoken word to written text without any errors and maintaining great quality in a short span of time.

Behind the scenes

The first step is converting the analogue waves that are created by a voice to digital signals that our computers can understand using analogue-to-digital converters (ADC). Now the digital form of sounds from ADC is cleaned and noises are eliminated alongside separating different soundtracks. The smallest segments of these signals are compared with the trained data sets using advanced machined learning algorithms and output text is derived from this activity.

Key applications of voice-based search

• Voice-based search is the most commonly used application of speech to text. It is widely used in banking, insurance, trade exchanges and even college libraries
• Speech recognition software can produce documents in less than half the time to type the same documents
• Hearing-impaired individuals can immediately understand the speech from the generated text without a time gap
• Voice recognition software can also be used in customer service to process regular phone orders, or in healthcare and law for documentation processes
• IT organisations use voice recognition software to improve internal communications and translate them in a data format that is easy to manage and search

Top leaders of voice recognition market

The voice recognition market is expected to reach to USD 31.82B by the year 2025 and large IT giants like Google, Amazon and IBM are already in the process of releasing their own APIs that convert speech to text in a short time. Also, in the open-source market, we can see that Simon, Kaldi, CMUSphinx are also popular to convert speech to text. It all depends on the user requirement to select the best technology from these available sources.

Still, a long way to go

It has been identified that more accuracy is needed, especially when there are glaring noises in the background. Separation of these signals needs to be refined to its best results. Along with this another area where a lot of research is going on is developing advanced artificial intelligence algorithms to build models on trained data.

In spite of few negative impacts like misuse of this in some privacy-related topics, the speech-to-text engines are giving a big edge to the human being in terms of solving some of the great issues and giving added customer satisfaction.

About the author

Dr P Balasubramanyam is an Associate Professor at GIM - Goa Institute of Management in the Big Data Analytics department with more than 20 years of working experience in IT and manufacturing domains. He led successfully many projects by applying Machine learning and AI techniques to complex manufacturing and connected industry domains. His primary research interest lies in the areas of supply chain management for manufacturing and predictive modelling for better decision making.

Note: The views expressed in this article are solely the author’s own and do not reflect/represent those of Shiksha