Startup Mantra: A successful note in giving solutions under speech analytics domain

Pune: Ever wondered why investigative agencies take time to solve complex cases? One of the reasons is lack of tools and technology to identify and validate the sample of any suspect’s speech, text, or video from the recorded sample and circumstantial evidence received. For example, a central investigative agency asks a suspect to do 300 signatures just to ensure that they can identify “skilled forgery”. After obtaining those signatures, they manually verify all the samples to arrive at any conclusion. Same is the case with any recorded audio evidence sample of a suspect, wherein the officers “listen” to the audio several times to conclude whether the evidence and recorded samples are of the same suspected individual.

A 69-year-old retired professor-turned-entrepreneur and founder of Anubhooti Solutions Shaila Dinkar Apte has designed and developed tools which helps in speaker (suspect’s voice sample) verification using word comparison, segregation of audio of different speakers, audio authentication (in case of tampered audio), emotion detection, noise cancellation, signature verification and document analysis and image authentication (in case of deep fake images).

Experiential teaching

A passionate researcher hailing from Thane, who was denied admission at a research institute for not fulfilling a certain criterion, Shaila pursued her dreams despite several domestic challenges. She started “coding” since 1980 when her visionary husband, an ex-employee of IIT Bombay, had brought a “personal computer” at their home. Shaila completed her mechanical engineering (ME) and PhD (electronics engineering) from Walchand College of Engineering at Sangli.

Shaila said, “As faculty of electronics and digital signal processing, I worked at Walchand College in Sangli for 28 years and at Rajarshi Shahu College of Engineering in Pune for 12 years since 2007. I had keenly taken interest in student projects and used to enjoy hard work for any innovative problem. Before retiring, at 62, I had registered my own proprietary firm named ‘Anubhooti Solutions’. I was guiding PhD candidates in my domain for topics like speech enhancement, writer verification, handwritten character recognition, speech disorder classification etc and even in other fields which I had a passion for. I have published five books with international publishers in signal processing area. This enriched my research experience with a team of 11 successful doctoral candidates. I guided over 60 ME candidates mostly working in speech processing domain with challenging research problems like emotional speech, audio signature, watermarking of audio, speaker verification and recognition using many different innovative ideas.”

“While helping students, I authored several books widely accepted by the industry. A Delhi-based startup founders approached me to help them in developing their voice-based product to generate stories in parents’ voice. I managed to make a prototype within six months, but since the founders did not have enough runway, the product development came to a halt. But this short stint gave me confidence about the industry outlook. I had rich research background, but no experience to run a company and commercial product development,” Shaila said.

Entrepreneurship

Shaila received another project from a Canada-based company for working on “emotion detection” and “speaker diarization” from voice samples. However, she was deceived by the firm which took the software code from her in exchange of just ₹2 lakh.

“I had never looked at the product from a money-making perspective and, hence, I had forgiven the Canada company. Meanwhile, another entrepreneur Suhas Kulkarni, the director of Lab Systems and Biotech (India) Private Limited, who worked for forensic labs and had seen the problems encountered by the agencies, approached me for developing some tools for audio analysis.

“Suhas and his friends had also worked in speech domain earlier. As part of their research, they came across my book and thought that the work will be done better by me. They requested me to generate a tool for audio analysis. I decided to do that as I had almost 40 years of experience in speech domain. Banking on my practical knowledge, I decided to do my best.

“Meanwhile, Covid-19 pandemic hit us and lockdown was announced. We utilised this period of about three years to develop the entire range of speaker diarization, speech lab, and speaker verification products market ready. We met only once before the lockdown, but afterwards, we had continued interaction for generating finished products. We started the work with a small team. Our only lack of experience was company operations and its related works,” Shaila said.

“There are around 200 forensic science labs (FSL) in India. These institutes train and educate in forensic domain, analyse audio clips grabbed by investigative agencies, and need tools for analysis. Currently, there are a few tools available for such analysis for forensic applications, but are non-user-friendly and difficult to operate. They are not found to obtain accurate results too. So, there is a need for good user-friendly and accurate tools. We took it as a challenge and developed six products for use of audio forensics. We believe that these made-in-India products will be useful for society,” she said.

Go-to-market

Anubhooti Solutions has collaborated with Lab Systems and Biotech (India) Private Limited and M/S Lab Systems (I) for marketing their products in India and across the globe. “We have focussed only on government sector so far and our customers are labs in forensic domain,” Shaila said.

Scaling up

Revealing her future development plans, Shaila said, “We are receiving requests from FSLs to generate a special dedicated system for detection of AI generated audios, images, and videos. Besides, the agencies are also in need of stress-detection feature along with emotion detection. We are working on this area and expect to launch a solution by the end of this financial year. We have Anubhooti signature verification, document analysis, and image authentication products in image analysis domain and video authentication will be launched soon.

“We have a permissible error margin of 2 to 3 per cent. Although we have automated the entire processes, the final call is always taken by an experienced analyst. Our tools will never replace them, but will help them in expediting the processes. The conclusions drawn on basis of reports or analysis from our tools are presented and admissible in court of law.”

No language barrier

Shaila said, “All tools developed by Anubhooti Solutions are language independent and can work for any Indian and foreign language. Thus, useful across the globe.”