Fueling AI Performance in Customer Service with Telephony Speech Data

Frost & Sullivan's "2023 China Intelligent Customer Service Market Report" reveals that the market size for intelligent customer service in China reached 6.7 billion yuan in 2022. With the continuous expansion of intelligent customer service applications, the market size is expected to grow to 18.1 billion yuan by 2027, demonstrating explosive growth potential.

During the 2023 World Artificial Intelligence Conference (WAIC), Yuan Xucong, Chief Analyst at Sullivan & Head of Panther Institute, pointed out that the emergence of large-scale AI models this year has provided a new breakthrough for the upgrading of intelligent customer service. Intelligent customer service will not be limited to traditional customer service domains but will also expand to marketing, sales, and other scenarios, bringing greater business value to enterprises.

Common forms of intelligent customer service include voice customer service, online customer service, chatbots, intelligent quality inspection, and digital human customer service. Among these, voice customer service serves as the fundamental customer interaction product in the customer service field.

In real-world environments, the application of speech recognition technology in telephone channel scenarios, including intelligent customer service, is increasing. However, telephone channels face diverse noise environments, including complex external conditions and linguistic diversity.

The complexity of external conditions encompasses noise, reverberation, echo, etc., which are distributed in different environments such as conference rooms, outdoors, and shopping malls. Similarly, the linguistic diversity poses a significant challenge, with different languages and even dozens of dialects within our own country.

With the rapid development of intelligent customer service technology, there is an urgent need for multilingual speech data recorded through telephone channels.

Challenges in collecting telephone channel speech data

-Speaker resources: Collecting foreign language speech requires expanding a sufficient number of foreign language speaker resources. For certain vertical domains like customer service, in addition to native speakers, communicative skills and the ability to guide conversations are required, preferably with relevant work experience.

-Corpus design: Designing natural dialogue corpora is highly challenging. Uniform recording scripts can sometimes restrict the speaker's creativity. Moreover, for specialized topics in certain domains, industry experts are needed to design the corpora.

-Quality control during the collection process: Controlling the recording environment is a primary requirement for success. If the recording room is spacious, it may result in reverberation, while surrounding noise sources can introduce various types of interference. If the recording environment is not well-controlled, almost all the speech data recorded in that environment will be deemed unqualified.

Nexdata Telephony Speech Data Solution

Nexdata possesses tens of thousands of hours of telephone channel speech data covering dozens of languages, including Chinese, English, Hindi, Thai, Italian, French, Spanish, and more. The topics covered range from daily life to vertical domains such as insurance, e-commerce, healthcare, and marketing.

Rich speaker resources

Over the years, through numerous overseas projects, Nexdata has accumulated over 200,000 crowdsourcing resources from dozens of countries, fully meeting the requirements for native speaker resources in different countries.

More Natural Dialogue Corpora

To ensure authentic natural dialogue styles for various everyday topics, we provide a topic list in advance. Recording participants can select multiple topics they are familiar with to ensure smooth and natural conversations, focusing on each topic and recording a dialogue segment.

Multiple Rounds of Quality Inspection

Nexdata has a mature and comprehensive project quality management system. We have strict and well-defined procedures for data quality acceptance to ensure that the delivered data fully meets customer requirements. Throughout the project implementation, Nexdata has established multi-round quality inspection systems for both data collection and annotation.

Having gained extensive experience in AI data services, Nexdata has established data security, compliance production, and data security management and monitoring systems, ensuring compliance in data collection.

Data Format

Nexdata's telephone channel speech dataset is recorded through a real telephone recording system in 8 kHz, 16-bit WAV format. Considering the use of both A-law and U-law algorithms in different regions worldwide, U-law is currently used by companies in North America and Japan, while A-law is used in Europe. Other regions use a mix of these two standards based on their country's situation. Nexdata adopts A-law/U-law PCM encoding formats for different language varieties during the collection process.

Annotation Content

All audio undergoes rigorous human transcription and quality inspection, including annotation of text content, start and end time points of valid sentences, and speaker identity labels, with an accuracy rate of over 95%.

Nexdata's multilingual telephone channel speech data can help enterprises enhance AI performance and build end-to-end services with high-quality datasets in intense market competition. The intelligent customer service market will witness explosive growth in the future, bringing more business opportunities and development space for enterprises.