Real-Time Voice AI Has a Low-Resource Language Problem: OpenAI Realtime and Gemini Live Fall Short for Azerbaijani
Key Takeaways
- ▸OpenAI Realtime comprehension works on clean speech but produces accent-distorted output unsuitable for professional interviews, with no model tuning or control available to fix phonology and stress patterns
- ▸Google Gemini Live unexpectedly handled comprehension and speech quality well but introduced latency severe enough to break conversational flow in real-time interactions
- ▸Developers targeting low-resource languages cannot rely on end-to-end vendor models and must build custom cascaded pipelines, taking ownership of latency budgets, turn-detection logic, and STT/TTS component integration
Summary
A technical evaluation of real-time voice AI systems reveals that both OpenAI's Realtime API and Google's Gemini Live have significant limitations when deployed in low-resource languages. The analysis, conducted while building an AI interview platform for Azerbaijani speakers, identifies distinct failure modes that don't appear in English demonstrations: comprehension degradation on accented speech, phonology errors in spoken output, and disruptive latency issues.
OpenAI Realtime achieved workable comprehension but produced Azerbaijani output with a distorted foreign accent unsuitable for professional applications. Gemini Live passed comprehension and accent quality tests but introduced response latencies that break conversational rhythm. The findings suggest a fundamental architectural problem: major end-to-end speech models optimize for high-resource languages, leaving developers in underserved markets with two unsatisfactory paths forward—accept degraded performance or build custom cascaded pipelines using tools like LiveKit, Pipecat, and Vapi.
This evaluation provides a decision framework for the growing number of companies building voice AI outside English-speaking markets, revealing how vendor performance claims valid in English can mask critical language-specific failures that only surface in production with real users.
- The evaluation reveals how vendor demos in English mask language-specific failure modes (comprehension degradation on accents/code-switching, asymmetric output quality, latency asymmetries) that only emerge in production with real candidates
Editorial Opinion
The gap between vendor-friendly English demos and production reality in low-resource languages exposes a blind spot in AI commercialization: major companies optimize for scale markets rather than linguistic diversity. While Realtime and Gemini Live represent genuine technical achievements, their failures in Azerbaijani highlight that real-world voice AI deployment is not yet a 'plug in an API and ship' problem. This evaluation underscores that for developers in underserved language markets, architectural flexibility and custom stack ownership remain unavoidable costs—and that the next wave of voice AI competitiveness may come from companies willing to solve this problem.
