labs · verba

Multilingual speech, made publishable.

Verba turns spoken notes into credible written output across English, 中文, español, français, and português. Raw transcript preserved. Ambiguity surfaced, not papered over. Drafts compared across providers, scored against a fixed rubric, and exportable as JSONL — all on the same Cloudflare edge that serves this site.

how it works
01

Transcribe

Drop audio. Whisper-style transcription with confidence-scored segments. Multilingual by default.

02

Polish

Rewrite, translate, or generate a prompt — three deliberate modes. Diff highlighted inline against the raw transcript.

03

Compare

Run the polished output through multiple providers in parallel. Side-by-side diff. Cost and latency captured per call.

04

Log eval

Score outputs with a fixed rubric. Persist to D1. Export the eval log as JSONL for downstream analysis.

Built end-to-end on Cloudflare. In daily use.