Anubodh/polyglot-voice-bench
0
Polyglot Voice Bench
A reproducible benchmark of voice AI providers on underrepresented languages: Hindi, Nepali, and English. 4 providers × 3 languages × 25 samples each, with 95% bootstrap confidence intervals.
Key findings:
- Deepgram Nova-3 leads on English (WER 0.086) and Hindi (WER 0.165)
- AssemblyAI Universal-3 Pro scores WER 1.008 on Hindi (CI [1.001–1.019]); every sample transcribed into a different script
- All providers fail on Nepali (WER 0.963–1.084); no provider signals unsupported language