Many organizations across the world today are increasingly using automatic transcription to help speed up processes. We’ve all heard of Alexa and Siri for taking our instructions, but what about e-commerce websites and client-facing organizations that need to automatically transcribe audio to text?
There is a growing demand for automatic transcription services, which means they need to be fast and produce accurate results. And with a widely spoken language like Hindi, the need to auto-transcribe has never been greater. In this article, we explore what automatic transcription is and what the challenges are to automatically transcribe audio to text in Hindi. Let’s get started!
What does automatic transcription mean?
An automatic transcription, broadly defined, is a term that refers to taking speech through an audio file, tokenizing that speech through a large language corpus, resulting in written text with high accuracy. This means that the spoken audio file is as accurately transcribed as possible. However, with Hindi, there are many challenges in this process. Below, we take a closer look at what some of these are.
Why will Hindi transcriptions probably have misinterpretations?
Despite the presence of automatic transcription services for Hindi and their continued development, automatic transcription of Hindi for commercial or non-commercial purposes poses certain challenges. Among these are some of the following:
- Hindi characters: in order to auto-transcribe audio from Hindi, the computer program needs to break down Hindi words into special characters. In Hindi, the alphabet script consists of vowels, consonants, and other characters. With regard to vowels, each vowel is represented by a separate symbol and there are 12 of them. However, the picture becomes more complicated because some consonants have an implicit vowel (matra) that is attached to the consonant. This, therefore, needs to be clearly distinguished by the software that is “reading” the sound file. In addition to vowels, consonants in Hindi are divided into different categories depending on the place and manner of their articulation. In particular, they are divided into five Vargs (groups) and nine non-Varg consonants. Some of these are nasal. Others constitute primary and secondary pairs. Some of these are unvoiced sounds whereas others are voiced sounds. And yet others are aspirated counterparts. Finally, with regard to the other characters, such as anuswar, visarga, chanderbindu – these can indicate nasal consonant sounds and each one will depend on the character that follows it. Therefore, this will determine whether the subsequent sound is nasal or not. As such, teaching a computer program to learn these distinct linguistic characteristics can prove challenging.
- Grapheme-to-phoneme (G2P) conversion: the second important challenge that arises comes with grapheme-to-phoneme (G2P) conversion in a computer language. This takes place when a written representation of a word or a combination of text forms is transcribed into a sound format.
- Schwa deletion: schwa deletion is a further challenge. This is because, in Hindi, some vowels at the beginning or end of certain words are completely omitted when spoken. Although, in written form, they are expressed fully.
- Compound words in Hindi: of course, Hindi is also characterized by compound words that are joined together to create meaning and context. As such, computer programs need to recognize this compound nature and ensure that they accurately auto-transcribe the spoken speech.
- Voice activity detection: a further challenge is voice activity detection. We all know that spoken language is not full of words only. Instead, it contains pauses and natural silence. In addition to this, there is also background noise that is picked up by computer systems, especially in a client-side usage of a mobile or web app when interacting with a computer interface. Thus, computer programs must be taught to recognize silences, pauses, as well as background noise and accurately tokenize these attributes to provide clear spacing between words.
- The need for an exceptionally large language corpus: when doing an automatic transcription for Hindi, there is also a need for an exceptionally large language corpus to ensure that when the computer program does an auto transcribe, it will have a large volume of data to use for more accurate transcription.
- Close collaborations are needed between linguists and computer scientists: of course, whether an organization requires an automatic video transcription or to auto-transcribe audio, there must be a close collaboration between linguists and computer scientists to ensure more accurate output.
- Implementation of speech recognition technology: and the final challenge on our list is the actual implementation of speech recognition technology when undertaking an automatic transcription. This can pose technical challenges for organizations that are not well-versed in the mechanics behind a technical implementation of an auto-transcribed audio file.
Case Studies
Despite the limited nature of the research that has so far been carried out in the field of automatic transcription – whether it’s to auto-transcribe audio or for an automatic video transcription – some scientists and authors have made great inroads into improving the automatic transcription process for Hindi using several different models and producing strong results with statistically significant outputs.
An example of this can be found in Kumar and Aggarwal’s work, which established that using their model for automatic transcription, the overall word accuracy and the word error rate of the system was 94.63% and 5.37%, respectively.
Apart from these authors, other works of Joshi and Kannan as well as Saha and Ramakrishnan in the field of auto-transcribing audio in Hindi, have also had statistically significant results.
Therefore, in the space where an organization needs to auto-transcribe for Hindi, positive strides have been made and further literature and studies need to be pursued for greater accuracy and better results for organizations.
Looking Ahead
To auto-transcribe in our day and age is a necessary part of doing business. It’s no longer about listening to voice recordings and manually typing them out. Instead, it’s about teaching computers to understand voice and then automatically transcribe it for business usage.
However, with the Hindi language, there are still many challenges that arise in the field of automatic transcription that’s free of errors. As such, further research is needed in this field to help organizations serve their customers better.