Reverie Sets A New Benchmark in Automated Speech Recognition (ASR) Accuracy for Indian Native Languages

Bengaluru, India, Friday 19th July 2024– Reverie Language Technologies, a pioneer in language technology solutions specializing in Indian languages, announces a significant advancement in its Automated Speech Recognition (ASR) model accuracy, validated through the esteemed Kathbath dataset curated by AI4Bharat. Reverie benchmarked their ASR output with Google and Microsoft for Indian English, Hindi, Gujarati, Marathi, Tamil, Telugu, Bengali, Kannada, Punjabi, Malayalam, and Odia. 

The Kathbath dataset, renowned for its comprehensive and diverse speech data, comprises an impressive 1,684 hours of meticulously human-labeled speech. This dataset spans 12 Indian languages and is contributed by 1,218 individuals hailing from 203 districts across the country. Such extensive linguistic diversity and regional representation make the Kathbath dataset a gold standard for evaluating ASR systems in the Indian context. 

Soham Bhattacharya, API Product Manager at Reverie, commented on the achievement “The established accuracy of Reverie’s generic ASR model is an outcome of the hard work and innovation of Reverie’s voice R&D division. By refining the fundamental algorithms and language models along with training the models on a huge quantity of data, we have set a new benchmark in the industry and thereby maintained the quality of transcriptions for our clients.” said Bhattacharya.

With over 15 years of experience, Reverie continues to lead the industry by achieving superior accuracy in ASR, surpassing standards with its latest innovations tailored for diverse sectors such as Banking, Finance, Insurance, e-commerce, Automotive, and Government.

Reverie’s Custom Model for Automated Speech Recognition

With Reverie’s Custom Speech-to-Text solutions, clients can fine-tune speech recognition models to perfectly match their specific needs.

For example, the phrase “Black shoe” may be expressed as “Kala Shoe” or as “Black Joota” reflecting regional linguistic nuances. ASR engineers work at creating multiple variations to encompass how individuals naturally search or speak, optimizing the model’s effectiveness across different linguistic contexts.

Reverie offers both online and offline models. The offline model is meticulously optimized to cater to a broad range of applications while efficiently managing memory resources. Reverie’s models also ensure maximum data security and privacy standards while seamlessly integrating into apps, websites, and systems. 

Acknowledging the limitations of offline environments, the engineers strategically reduce the size of these models without sacrificing performance. This approach allows Reverie to support multiple use cases concurrently, providing accessible and responsive solutions for its clientele. The design ensures comprehensive functionality, even in settings where continuous internet connectivity may be unreliable.

ASR Live Use Cases for e-commerce & Automobile

Reverie’s commitment to innovation can be validated by its collaboration with a global sporting goods retailer, which recently launched Multilingual Voice Search on its e-commerce platform. This initiative, supporting Indian English, Hindi and Tamil, has significantly enhanced customer engagement, handling over 4000 voice searches per day within its first week of launch, motivating the client to add three more Indian languages to the platform.  

Additionally working with the automotive sector, Reverie has revolutionized in-car voice assistants with a robust multilingual offline ASR model. Addressing the unique challenges of automotive environments, such as engine noise and multiple in-car passenger conversations, Reverie’s solution ensures accurate and reliable performance. It allows users to control settings, access internet information, and perform calculations using simple voice commands. Some of the key features include automatic speech recognition, natural language understanding, and text-to-speech, complemented by an offline wake word system – ‘Hello Car’. Implemented in a prominent Indian automobile brand, this technology enhances convenience and safety, transforming the driving experience.

About Reverie

Reverie is India’s 1st AI-powered language technology platform that offers next-generation text, voice, and video localisation solutions. Reverie commits to creating language equality in the digital world and empowers government agencies and private enterprises to reach billions of people across the country in their native languages. Leveraging advanced technologies, such as Neural Machine Translation (NMT), Natural Language Understanding (NLU), and Machine Translation (ML), Reverie is ahead of the curve in enabling digital initiatives across India. 

For more information on Reverie’s ASR solutions and custom Model STT/ASR technology, visit Reverie.