Case Studies > How We Help Train Your Machines to Understand Multilingual Voice Search

How We Help Train Your Machines to Understand Multilingual Voice Search

Company Size

1,000+

Country

Australia
United Kingdom
United States

Product

Lionbridge Speech Collection Platform
Lionbridge Speech Testing Services
Lionbridge Voice Recording Booths

Tech Stack

Speech Recognition Algorithms
Natural Language Processing (NLP)
Voice Recording Studios

Implementation Scale

Enterprise-wide Deployment

Impact Metrics

Brand Awareness
Customer Satisfaction
Digital Expertise

Technology Category

Analytics & Modeling - Machine Learning
Analytics & Modeling - Natural Language Processing (NLP)
Application Infrastructure & Middleware - Data Exchange & Integration

Applicable Industries

Consumer Goods
Software
Telecommunications

Applicable Functions

Business Operation
Product Research & Development
Quality Assurance

Use Cases

Speech Recognition

Services

Data Science Services
System Integration
Testing & Certification

About The Customer

One of the world’s largest software and technology enterprises, our client produces a multilingual search engine that delivers text, image, video, and map results for text- and voice-based searches. The client’s voice-activated personal assistant app assists users by answering questions, making recommendations, and performing actions. Providing global users with accurate results to their queries is of paramount importance to our client, which strives to generate accurate results and actions based on voice input using natural language in most spoken languages.

The Challenge

Around the world, the popularity of voice search is only growing. As our client seeks to provide fast and accurate search and voice assistance services to users of every major language worldwide, it must understand a variety of voice-activated queries in multiple languages. Global speakers of major languages expect voice-activated apps and search engines to fully and correctly understand voice queries, regardless of the type of voice, regional accent, age, gender, device used, or surrounding environment. With billions of global users and the complexities of processing natural language, fully and correctly recognizing any user’s speech in any language is our client’s complex challenge.

The Solution

To ensure its technology can generate the correct output to a given query, our client requires large quantities of speech samples across voice types, genders, languages, accents, and environments. These speech samples serve, effectively, to train its speech recognition algorithms. Our client partnered with Lionbridge for a one-stop, end-to-end solution that catered to each step of the speech creation, development, and testing process. Lionbridge utilized its global network of voice recording booths and studios to deliver multi-lingual, studio-quality voice recordings of selected content. Lionbridge tapped into its global crowd of 400,000+ members to collect the speech data required by the client. Using our streamlined platform that supports participant registration, evaluation, and workflow management, Lionbridge collected over 20 hours of speech samples across 30 languages and from speakers aged 6 to 75. Lionbridge captured 240 hours of high-quality ambient noise from various environments across dozens of countries and languages, including recordings in cafes, cars, pubs, restaurants, malls, schools, homes, offices, streets, train stations, and airports.

Operational Impact

Lionbridge partnered with our client to amass a data set that spanned 30+ languages and dialects, regions, voice types, and ambient noises.
With the data collected, our client can develop, calibrate, and improve its speech recognition algorithms, testing them with real speech from around the world, in a variety of real and simulated scenarios.
This enables our client to generate more accurate search results and correct actions based on user commands, improving the user experience, and strengthening brand loyalty worldwide.

Quantitative Benefit