A global voice-AI company building speech recognition for consumer devices across Asia-Pacific and African language markets

This is a Lifewood Data Technology case study in Multilingual speech + LLM — an engagement delivered for a global voice-AI company building speech recognition for consumer devices across Asia-Pacific and African language markets across China · Asia-Pacific. 14,000 hours of speech across 11 languages and 8 countries, cutting word error rate 92%

A voice assistant that worked in 15 languages failed everywhere it needed to grow

The client's voice assistant worked well in 15 major languages but failed in emerging markets where hundreds of millions of potential users spoke languages with minimal digital speech data available. Off-the-shelf speech datasets did not exist for the target languages, and the client could not ethically scrape speech data at scale. A ground-up collection effort was required across geographies where the client had no operational footprint. The constraint that made it harder was representativeness. Speakers had to span diverse ages, genders, accents and recording environments, so the finished voice AI would work for the full population rather than an urban subset. A dataset collected conveniently in one place reproduces that place's demographics no matter how large it grows — which is precisely the failure mode crowdsourcing platforms produce, since they skew toward urban, educated, high-resource-language populations and systematically miss dialect and register variation.

Field operations in eight countries, recruiting in-community rather than online

Lifewood activated its field operations in 8 African and Southeast Asian countries, recruiting 6,200+ native speakers across rural and urban communities. Contributors were ethically compensated above local fair-wage benchmarks with full informed-consent documentation. Collection covered scripted prompts, spontaneous conversation, and domain-specific commands across commerce, navigation and media, tailored to the client's assistant use cases. 1. In-community recruitment. Contributors are recruited from the language community itself through field operations, not from crowdsourcing platforms — which is what makes authentic dialect, register and accent coverage possible at all. 2. Demographic balancing by design. Speaker panels were balanced across age, gender, accent and recording environment as a recruitment specification rather than as a post-hoc filter. 3. Environmental diversity sampling. Recording conditions were varied deliberately so the model learns the speech rather than the studio. 4. Phonetic transcription and quality control. Quality control included phonetic transcription, speaker demographic balancing and environmental noise diversity sampling, under Lifewood's 95%+ accuracy SLA and dual-layer human-in-the-loop review. Lifewood's language coverage behind this programme spans 50+ languages reaching more than 90% of the global population, including Swahili, Wolof, Hausa, Amharic, Tigrinya, Yoruba, Zulu, Shona, Lingala and Somali across Africa, and Tagalog, Cebuano, Ilokano, Waray, Khmer, Tok Pisin, Tetum, Fijian and Samoan across Southeast Asia and the Pacific.

Word error rate fell 92% and eleven new market languages shipped within nine months

The programme delivered 14,000 hours of speech across 11 languages from 6,200+ unique speakers in 8 countries. On the client side, word error rate for target languages fell 92% against the pre-programme baseline, and 11 new market languages went live in the assistant within nine months of programme start — against a working base of 15 major languages before the engagement began.

Related Lifewood services

• low-resource speech data collection
• multilingual training data collection
• enterprise LLM training data