The voice-to-text transcription market is growing, with applications spanning the most diverse industries. But what are the killer features to focus attention on in the next two or three years? We proposed some of them.
Accuracy Of Speech Recognition
For audio transcriptions in the most popular languages (e.g., English), accuracy has been achieved in some scenarios close to that of humans, with values close to 95% (therefore with WER – word error rate – of 5 %). Therefore, those involved in speech recognition technology will have to bring the same standards of accuracy to other languages, making sure to reach levels that are more than satisfactory for increasingly multinational and multilingual realities (companies and, consequently, final customers).
In addition to these capabilities, players will have to include in the offer solutions aimed at improving the quality of the output provided to their customers, ranging from the identification of spoken languages to speaker diarization. (more on this later), ensuring that the promised levels of accuracy truly materialize as they apply to the real world. An example that is often taken for granted is the ability to provide quality transcription output in noisy environments, spontaneous speech conversations, or audio recorded on low-quality devices.
Identify Who Says What
Speaker diarization is used to identify the voices of individuals in audio / multimedia files recorded on a single channel: the “unique speakers” are identified by assigning a label to each of them and associating it with the corresponding text portions in the transcription.
This activity represents a real challenge for automated systems: a single speaker can vary tone and way of speaking according to mood, hesitation, the emphasis he wants to give to words, surrounding noise, and many other variables, therefore bringing all its nuances back to a single label, differentiating it from the others, is not as obvious as it might seem.
Automatic Identification Of The Spoken Language
Automating the step of identifying the spoken language in real-time before starting the transcription process (an activity that would otherwise occur with manual selection of the correct language pack ) allows companies to simplify the management of business processes in multilingual contexts, avoiding assets whose voice is lost or is available too late.
Customization Of Acoustic Models And Language
The availability of proprietary technologies is the driving force for obtaining the best automatic transcription performance. The adaptation of acoustic and language models, specific to the reference context, allows you to transcribe audio into the text from a wide range of inputs (telephone, broadcast, …), obtaining high-quality standards, as well as overcoming the obstacles of voice recognition due to particular acoustic environments and specific domain terminologies (e.g., names of structures, products, brands, acronyms used by the customer).
The ability to fine-tune these models makes it possible to guarantee more adequate and precise outputs, unlike the adoption of general-purpose systems. But the path must be refined, also through closer collaboration between users and suppliers, starting with the sharing of data and up to the achievement (progressive and incremental) of truly effective outputs.
Extension Of The Capabilities Of Virtual Assistants
Given the great attention paid to the world of virtual assistants and their growing use on smartphones and other devices, it is essential to increase accuracy even in particular contexts and application scenarios (e.g., support for blocking credit cards, booking medical examinations, …). Consumers expect their virtual assistants to understand them, regardless of their accent, dialect, or language, even in the case of sentences that are not always easily contextualized.
Speech-To-Text And Machine Translation In The Target Language
A typical need of companies operating globally is to use a common, unique, and recognizable language that represents them, regardless of the country in which the communications (institutional or internal to the company itself) will be conveyed. In this scenario, it is necessary to have advanced tools that allow immediate transcription from the speaker’s language to that of the listener. The solutions for multilingual communication must then allow high levels of accuracy and reduced latency.