AI article
The Verb Comes Last: Measuring Why Japanese-English Live Translation Waits
Community description: Built an all-local streaming JA-EN speech translator and measured the verb-final wait cost: EN-JA prefix translations survive ~13x worse than JA-EN at matched prefixes (same backend).
Dev.to | Oct 20, 2026 | Raihan
Automated excerpt
On 200 utterances per direction: ASR CER 0. 091 (Japanese) / WER 0. 104 (English); MT BLEU barely drops when fed ASR output instead of clean text (10. 6 vs 18. 1 ja→en; 24. 1 vs 24. 3 en→ja). Per-stage GPU cost: ASR ~0. 35s, opus MT 1. 39s, Qwen MT 0. 38s, TTS ~0. 2s. A lockstep streaming sim (VAD segments → ASR → MT → TTS) on 20 utterances per direction: first-audio lag p50 1. 86s ja→en, 1. 26s en→ja (p95 3. 32s / 2. 22s).
Selected automatically from source text; not independently written or fact-checked. Read the original for full context.