Smart use of crowdsourcing in the implementation of Voice-driven Solutions

One of the best scenarios in offering natural language and voice-enabled services is having freedom to construct, tweak, and optimise bespoke products that perfectly meet the needs of one’s business. Achieving the highest possible level of accuracy in speech recognition and voice biometrics is key to ensuring good performance from the start.

This means in most cases, that the solution should use the model trained on the voices of your existing customers, which creates personal data protection challenges, especially if this data needs to be transferred to a vendor. Even if the customer data is pseudonymised, it will still be regarded as personal data.

Using generic models offered by the large companies does not deliver the same level of accuracy as bespoke solutions, no matter how hard you try. It is also not a very good idea to share your customer data with these large companies in any form, thereby increasing their competitiveness and compromising your own.

But there is a way around it – crowdsourcing tools that use voices from the same country and locality, even voices of the people speaking the same dialect, if that is the challenge (e.g. Swiss German speaking cantons of Switzerland for an SBB voice-driven app – refer to an overview of this example in the chart below).

Spitch’s Lingware Suite makes it possible to rapidly collect targeted spoken language data for a specific application or business case, and support your own high quality spoken language processing services. The Lingware Suite can either be used in-cloud to tap into the power of a crowdsourcing community, or delivered on-premise to mobilise the client’s own staff and customer data for specific tasks, saving time and significant resources.

Among others, it can deliver the following benefits:

• Voice & Text Data Collection (Cloud-based).
  Create and deploy targeted crowd-sourced voice and text data collection applications. Bootstrap new applications and help guarantee improved performance by directly targeting your prospective application domain, vocabulary, and clientele.
• Data Annotation & Transcription (Cloud-based).
  Create and deploy crowd-sourced transcription and annotation tasks to transform your raw data into knowledge resources enabling your business to become intelligently client-centric.
• SLT Application Builder.
  Build custom Spoken Language Technology (SLT) applications that combine Automatic Speech Recognition (ASR) - GRXML and LVCSR, Statistical Semantic Models (SSM), Speaker Verification (SV) and more. Rapidly test, revise and deploy new applications
• Semantic Modeling.
  Train and configure Statistical Semantic Models (SSM) to use as standalone text classifiers, or as a Natural Language Understanding (NLU) module layered on top of an SLT application.
• Speech-to-text & backend.
  Prepare and configure customised speech-to-text and semantic interpretation resources for use in multi-channel applications or direct API access. Add speaker identification and verification services using full-featured and scalable core biometric solution
• Administration & Monitoring.
  Benefit from a system user account management. There are different access levels for different users and integrated applications.
• Management & Dashboards and reporting.
  Create dashboards easily for call-center solutions. Save work-time with an Agent’s KPI builder. Notification and reporting system can alert the supervisor when bad sentiment is detected or some index is out of bounds. Fraudsters detection dashboard is also available.
• Telephony Integration.
  Quickly and easily plug your application into the telephony network. We are fully compatible with all types of PBXs (AVAYA, Genesys, FreeSWITCH, Cisco etc.). Add a high-quality recording system that can store records in lossless formats for further analysis.

Doing data collection and annotation for models training can be a cumbersome business, as you would quickly discover when trying to increase the accuracy of speech recognition solutions, while adhering to customer data protection and privacy regulations. Providing proper lingware tools for models training by clients themselves or their partners we can help get the same revenue with less investment.

Crowdsourcing opportunities:

• Crowdsourcing platforms are seamlessly integrated for data collection and annotation tasks;
• Crowdsourcing annotation is cloud-based;
• Pre-annotation sensitive data reduction is used to ensure regulatory compliance and protect clients’ data.

Spitch’s Lingware Suite crowdsourcing instruments can be delivered in two ways:

In cloud This product version is highly integrated in the AWS services. It has all the features of an on-premise version. This approach is ideal for crowdsourcing.

On-premise Delivered as the OVA file, which can be deployed in the Virtualization Environment of the enterprise client. On-premise version is most suitable for using customer data and collecting data for annotation internally e.g. from staff-members in large organisations.

To summarise: Lingware Suite delivers at least three very tangible advantages:

1. High speed – decreasing time-to-market and helping bring the product from the idea to the customer in a very short time.
2. Cost saving – providing easy-to-use tools for models training by clients themselves or their partners we can help cut costs that are inevitably higher in all other scenarios offered by vendors, including outsourcing options.
3. Security and compliance – in most enterprise deployment cases, one cannot use client’s customer data for model improvement due to confidentiality, regulatory compliance constraints, or for other reasons. In this situation, the Lingware Suite offers a unique set of tools to increase the accuracy of speech recognition by collecting targeted and client-specific data for supervised machine learning in accordance with the regulatory framework.

An additional benefit is that it also becomes easier to improve AI quality through constant additional training as part of support & maintenance. This kind of supervised machine learning application allows to fix specific issues that arise with one customer making sure they do not reappear with all other customers.